bigtime

analyze_memory_profile.py

@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+"""
+Analyze memory profile JSON files to identify OOM causes
+"""
+
+import json
+import glob
+import os
+import sys
+from pathlib import Path
+
+def analyze_memory_files():
+    """Analyze partial memory profile files"""
+    
+    # Get all partial files in order
+    files = sorted(glob.glob('memory_profile_partial_*.json'))
+    
+    if not files:
+        print("❌ No memory profile files found!")
+        print("Expected files like: memory_profile_partial_0.json")
+        return
+        
+    print(f"🔍 Found {len(files)} memory profile files")
+    print("=" * 60)
+    
+    peak_memory = 0
+    peak_vram = 0
+    critical_points = []
+    all_checkpoints = []
+    
+    for i, file in enumerate(files):
+        try:
+            with open(file, 'r') as f:
+                data = json.load(f)
+            
+            timeline = data.get('timeline', [])
+            if not timeline:
+                continue
+                
+            # Find peaks in this file
+            file_peak_rss = max([d['rss_gb'] for d in timeline])
+            file_peak_vram = max([d['vram_gb'] for d in timeline])
+            
+            if file_peak_rss > peak_memory:
+                peak_memory = file_peak_rss
+            if file_peak_vram > peak_vram:
+                peak_vram = file_peak_vram
+                
+            # Find memory growth spikes (>3GB increase)
+            for j in range(1, len(timeline)):
+                prev_rss = timeline[j-1]['rss_gb']
+                curr_rss = timeline[j]['rss_gb']
+                growth = curr_rss - prev_rss
+                
+                if growth > 3.0:  # >3GB growth spike
+                    checkpoint = timeline[j].get('checkpoint', f'sample_{j}')
+                    critical_points.append({
+                        'file': file,
+                        'file_index': i,
+                        'sample': j,
+                        'timestamp': timeline[j]['timestamp'],
+                        'rss_gb': curr_rss,
+                        'vram_gb': timeline[j]['vram_gb'],
+                        'growth_gb': growth,
+                        'checkpoint': checkpoint
+                    })
+                    
+            # Collect all checkpoints
+            checkpoints = [d for d in timeline if 'checkpoint' in d]
+            for cp in checkpoints:
+                cp['file'] = file
+                cp['file_index'] = i
+                all_checkpoints.append(cp)
+                
+            # Show progress for this file
+            if timeline:
+                start_rss = timeline[0]['rss_gb']
+                end_rss = timeline[-1]['rss_gb']
+                growth = end_rss - start_rss
+                samples = len(timeline)
+                
+                print(f"📊 File {i+1:2d}: {start_rss:5.1f}GB → {end_rss:5.1f}GB "
+                      f"(+{growth:4.1f}GB) [{samples:3d} samples]")
+                
+                # Show significant checkpoints from this file
+                if checkpoints:
+                    for cp in checkpoints:
+                        print(f"    📍 {cp['checkpoint']}: {cp['rss_gb']:.1f}GB")
+                        
+        except Exception as e:
+            print(f"❌ Error reading {file}: {e}")
+    
+    print("\n" + "=" * 60)
+    print("🎯 ANALYSIS SUMMARY")
+    print("=" * 60)
+    
+    print(f"📈 Peak Memory: {peak_memory:.1f} GB")
+    print(f"🎮 Peak VRAM: {peak_vram:.1f} GB")
+    print(f"⚡ Growth Spikes: {len(critical_points)} events >3GB")
+    
+    if critical_points:
+        print(f"\n💥 MEMORY GROWTH SPIKES (>3GB):")
+        print("   Location                           Growth   Total   VRAM")
+        print("   " + "-" * 55)
+        
+        for point in critical_points:
+            location = point['checkpoint'][:30].ljust(30)
+            print(f"   {location} +{point['growth_gb']:4.1f}GB → {point['rss_gb']:5.1f}GB  {point['vram_gb']:4.1f}GB")
+    
+    if all_checkpoints:
+        print(f"\n📍 CHECKPOINT PROGRESSION:")
+        print("   Checkpoint                         Memory   VRAM    File")
+        print("   " + "-" * 55)
+        
+        for cp in all_checkpoints:
+            checkpoint = cp['checkpoint'][:30].ljust(30)
+            file_num = cp['file_index'] + 1
+            print(f"   {checkpoint} {cp['rss_gb']:5.1f}GB  {cp['vram_gb']:4.1f}GB  #{file_num}")
+    
+    # Memory growth analysis
+    if len(all_checkpoints) > 1:
+        print(f"\n📊 MEMORY GROWTH ANALYSIS:")
+        
+        # Find the biggest memory jumps between checkpoints
+        big_jumps = []
+        for i in range(1, len(all_checkpoints)):
+            prev_cp = all_checkpoints[i-1]
+            curr_cp = all_checkpoints[i]
+            
+            growth = curr_cp['rss_gb'] - prev_cp['rss_gb']
+            if growth > 2.0:  # >2GB jump
+                big_jumps.append({
+                    'from': prev_cp['checkpoint'],
+                    'to': curr_cp['checkpoint'],
+                    'growth': growth,
+                    'from_memory': prev_cp['rss_gb'],
+                    'to_memory': curr_cp['rss_gb']
+                })
+        
+        if big_jumps:
+            print("   Major jumps (>2GB):")
+            for jump in big_jumps:
+                print(f"   {jump['from']} → {jump['to']}: "
+                      f"+{jump['growth']:.1f}GB ({jump['from_memory']:.1f}→{jump['to_memory']:.1f}GB)")
+        else:
+            print("   ✅ No major memory jumps detected")
+    
+    # Diagnosis
+    print(f"\n🔬 DIAGNOSIS:")
+    
+    if peak_memory > 400:
+        print("   🔴 CRITICAL: Memory usage exceeded 400GB")
+        print("   💡 Recommendation: Reduce chunk_size to 200-300 frames")
+    elif peak_memory > 200:
+        print("   🟡 HIGH: Memory usage over 200GB")
+        print("   💡 Recommendation: Reduce chunk_size to 400 frames")
+    else:
+        print("   🟢 MODERATE: Memory usage under 200GB")
+    
+    if critical_points:
+        # Find most common growth spike locations
+        spike_locations = {}
+        for point in critical_points:
+            location = point['checkpoint']
+            spike_locations[location] = spike_locations.get(location, 0) + 1
+        
+        print("\n   🎯 Most problematic locations:")
+        for location, count in sorted(spike_locations.items(), key=lambda x: x[1], reverse=True)[:3]:
+            print(f"      {location}: {count} spikes")
+    
+    print(f"\n💡 NEXT STEPS:")
+    if 'merge' in str(critical_points).lower():
+        print("   1. Chunk merging still causing memory accumulation")
+        print("   2. Check if streaming merge is actually being used")
+        print("   3. Verify chunk files are being deleted immediately")
+    elif 'propagation' in str(critical_points).lower():
+        print("   1. SAM2 propagation using too much memory")
+        print("   2. Reduce chunk_size further (try 300 frames)")
+        print("   3. Enable more aggressive frame release")
+    else:
+        print("   1. Review the checkpoint progression above")
+        print("   2. Focus on locations with biggest memory spikes")
+        print("   3. Consider reducing chunk_size if spikes are large")
+
+def main():
+    print("🔍 MEMORY PROFILE ANALYZER")
+    print("Analyzing memory profile files for OOM causes...")
+    print()
+    
+    analyze_memory_files()
+
+if __name__ == "__main__":
+    main()

debug_memory_leak.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+"""
+Debug memory leak between chunks - track exactly where memory accumulates
+"""
+
+import psutil
+import gc
+from pathlib import Path
+import sys
+
+def detailed_memory_check(label):
+    """Get detailed memory info"""
+    process = psutil.Process()
+    memory_info = process.memory_info()
+    
+    rss_gb = memory_info.rss / (1024**3)
+    vms_gb = memory_info.vms / (1024**3)
+    
+    # System memory
+    sys_memory = psutil.virtual_memory()
+    available_gb = sys_memory.available / (1024**3)
+    
+    print(f"🔍 {label}:")
+    print(f"   RSS: {rss_gb:.2f} GB (physical memory)")
+    print(f"   VMS: {vms_gb:.2f} GB (virtual memory)")
+    print(f"   Available: {available_gb:.2f} GB")
+    
+    return rss_gb
+
+def simulate_chunk_processing():
+    """Simulate the chunk processing to see where memory accumulates"""
+    
+    print("🚀 SIMULATING CHUNK PROCESSING TO FIND MEMORY LEAK")
+    print("=" * 60)
+    
+    base_memory = detailed_memory_check("0. Baseline")
+    
+    # Step 1: Import everything (with lazy loading)
+    print("\n📦 Step 1: Imports")
+    from vr180_matting.config import VR180Config
+    from vr180_matting.vr180_processor import VR180Processor
+    
+    import_memory = detailed_memory_check("1. After imports")
+    import_growth = import_memory - base_memory
+    print(f"   Growth: +{import_growth:.2f} GB")
+    
+    # Step 2: Load config
+    print("\n⚙️  Step 2: Config loading")
+    config = VR180Config.from_yaml('config.yaml')
+    config_memory = detailed_memory_check("2. After config load")
+    config_growth = config_memory - import_memory
+    print(f"   Growth: +{config_growth:.2f} GB")
+    
+    # Step 3: Initialize processor (models still lazy)
+    print("\n🏗️  Step 3: Processor initialization")
+    processor = VR180Processor(config)
+    processor_memory = detailed_memory_check("3. After processor init")
+    processor_growth = processor_memory - config_memory
+    print(f"   Growth: +{processor_growth:.2f} GB")
+    
+    # Step 4: Load video info (lightweight)
+    print("\n🎬 Step 4: Video info loading")
+    try:
+        video_info = processor.load_video_info(config.input.video_path)
+        print(f"   Video: {video_info.get('width', 'unknown')}x{video_info.get('height', 'unknown')}, "
+              f"{video_info.get('total_frames', 'unknown')} frames")
+    except Exception as e:
+        print(f"   Warning: Could not load video info: {e}")
+    
+    video_info_memory = detailed_memory_check("4. After video info")
+    video_info_growth = video_info_memory - processor_memory
+    print(f"   Growth: +{video_info_growth:.2f} GB")
+    
+    # Step 5: Simulate chunk 0 processing (this is where models actually load)
+    print("\n🔄 Step 5: Simulating chunk 0 processing...")
+    
+    # This is where the real memory usage starts
+    print("   Loading first 10 frames to trigger model loading...")
+    try:
+        # Read a small number of frames to trigger model loading
+        frames = processor.read_video_frames(
+            config.input.video_path,
+            start_frame=0,
+            num_frames=10,  # Just 10 frames to trigger model loading
+            scale_factor=config.processing.scale_factor
+        )
+        
+        frames_memory = detailed_memory_check("5a. After reading 10 frames")
+        frames_growth = frames_memory - video_info_memory
+        print(f"   10 frames growth: +{frames_growth:.2f} GB")
+        
+        # Free frames
+        del frames
+        gc.collect()
+        
+        after_free_memory = detailed_memory_check("5b. After freeing 10 frames")
+        free_improvement = frames_memory - after_free_memory
+        print(f"   Memory freed: -{free_improvement:.2f} GB")
+        
+    except Exception as e:
+        print(f"   Could not simulate frame loading: {e}")
+        after_free_memory = video_info_memory
+    
+    print(f"\n📊 MEMORY ANALYSIS:")
+    print(f"   Baseline → Final: {base_memory:.2f}GB → {after_free_memory:.2f}GB")
+    print(f"   Total growth: +{after_free_memory - base_memory:.2f}GB")
+    
+    if after_free_memory - base_memory > 10:
+        print(f"   🔴 HIGH: Memory growth > 10GB before any real processing")
+        print(f"   💡 This suggests model loading is using too much memory")
+    elif after_free_memory - base_memory > 5:
+        print(f"   🟡 MODERATE: Memory growth 5-10GB")
+        print(f"   💡 Normal for model loading, but monitor chunk processing")
+    else:
+        print(f"   🟢 GOOD: Memory growth < 5GB")
+        print(f"   💡 Initialization memory usage is reasonable")
+    
+    print(f"\n🎯 KEY INSIGHTS:")
+    if import_growth > 1:
+        print(f"   - Import growth: {import_growth:.2f}GB (fixed with lazy loading)")
+    if processor_growth > 10:
+        print(f"   - Processor init: {processor_growth:.2f}GB (investigate model pre-loading)")
+    
+    print(f"\n💡 RECOMMENDATIONS:")
+    if after_free_memory - base_memory > 15:
+        print(f"   1. Reduce chunk_size to 200-300 frames")
+        print(f"   2. Use smaller models (yolov8n instead of yolov8m)")
+        print(f"   3. Enable FP16 mode for SAM2")
+    elif after_free_memory - base_memory > 8:
+        print(f"   1. Monitor chunk processing carefully")
+        print(f"   2. Use streaming merge (should be automatic)")
+        print(f"   3. Current settings may be acceptable")
+    else:
+        print(f"   1. Settings look good for initialization")
+        print(f"   2. Focus on chunk processing memory leaks")
+
+def main():
+    if len(sys.argv) != 2:
+        print("Usage: python debug_memory_leak.py <config.yaml>")
+        print("This simulates initialization to find memory leaks")
+        sys.exit(1)
+    
+    config_path = sys.argv[1]
+    if not Path(config_path).exists():
+        print(f"Config file not found: {config_path}")
+        sys.exit(1)
+    
+    simulate_chunk_processing()
+
+if __name__ == "__main__":
+    main()

memory_profiler_script.py

@@ -0,0 +1,249 @@
+#!/usr/bin/env python3
+"""
+Memory profiling script for VR180 matting pipeline
+Tracks memory usage during processing to identify leaks
+"""
+
+import sys
+import time
+import psutil
+import tracemalloc
+import subprocess
+import gc
+from pathlib import Path
+from typing import Dict, List, Tuple
+import threading
+import json
+
+class MemoryProfiler:
+    def __init__(self, output_file: str = "memory_profile.json"):
+        self.output_file = output_file
+        self.data = []
+        self.process = psutil.Process()
+        self.running = False
+        self.thread = None
+        self.checkpoint_counter = 0
+        
+    def start_monitoring(self, interval: float = 1.0):
+        """Start continuous memory monitoring"""
+        tracemalloc.start()
+        self.running = True
+        self.thread = threading.Thread(target=self._monitor_loop, args=(interval,))
+        self.thread.daemon = True
+        self.thread.start()
+        print(f"🔍 Memory monitoring started (interval: {interval}s)")
+        
+    def stop_monitoring(self):
+        """Stop monitoring and save results"""
+        self.running = False
+        if self.thread:
+            self.thread.join()
+        
+        # Get tracemalloc snapshot
+        snapshot = tracemalloc.take_snapshot()
+        top_stats = snapshot.statistics('lineno')
+        
+        # Save detailed results
+        results = {
+            'timeline': self.data,
+            'top_memory_allocations': [
+                {
+                    'file': stat.traceback.format()[0],
+                    'size_mb': stat.size / 1024 / 1024,
+                    'count': stat.count
+                }
+                for stat in top_stats[:20]  # Top 20 allocations
+            ],
+            'summary': {
+                'peak_rss_gb': max([d['rss_gb'] for d in self.data]) if self.data else 0,
+                'peak_vram_gb': max([d['vram_gb'] for d in self.data]) if self.data else 0,
+                'total_samples': len(self.data)
+            }
+        }
+        
+        with open(self.output_file, 'w') as f:
+            json.dump(results, f, indent=2)
+        
+        tracemalloc.stop()
+        print(f"📊 Memory profile saved to {self.output_file}")
+        
+    def _monitor_loop(self, interval: float):
+        """Continuous monitoring loop"""
+        while self.running:
+            try:
+                # System memory
+                memory_info = self.process.memory_info()
+                rss_gb = memory_info.rss / (1024**3)
+                
+                # System-wide memory
+                sys_memory = psutil.virtual_memory()
+                sys_used_gb = (sys_memory.total - sys_memory.available) / (1024**3)
+                sys_available_gb = sys_memory.available / (1024**3)
+                
+                # GPU memory (if available)
+                vram_gb = 0
+                vram_free_gb = 0
+                try:
+                    result = subprocess.run(['nvidia-smi', '--query-gpu=memory.used,memory.free', 
+                                           '--format=csv,noheader,nounits'], 
+                                          capture_output=True, text=True, timeout=5)
+                    if result.returncode == 0:
+                        lines = result.stdout.strip().split('\n')
+                        if lines and lines[0]:
+                            used, free = lines[0].split(', ')
+                            vram_gb = float(used) / 1024
+                            vram_free_gb = float(free) / 1024
+                except Exception:
+                    pass
+                
+                # Tracemalloc current usage
+                try:
+                    current, peak = tracemalloc.get_traced_memory()
+                    traced_mb = current / (1024**2)
+                except Exception:
+                    traced_mb = 0
+                
+                data_point = {
+                    'timestamp': time.time(),
+                    'rss_gb': rss_gb,
+                    'vram_gb': vram_gb,
+                    'vram_free_gb': vram_free_gb,
+                    'sys_used_gb': sys_used_gb,
+                    'sys_available_gb': sys_available_gb,
+                    'traced_mb': traced_mb
+                }
+                
+                self.data.append(data_point)
+                
+                # Print periodic updates and save partial data
+                if len(self.data) % 10 == 0:  # Every 10 samples
+                    print(f"🔍 Memory: RSS={rss_gb:.2f}GB, VRAM={vram_gb:.2f}GB, Sys={sys_used_gb:.1f}GB")
+                    
+                    # Save partial data every 30 samples in case of crash
+                    if len(self.data) % 30 == 0:
+                        self._save_partial_data()
+                    
+            except Exception as e:
+                print(f"Monitoring error: {e}")
+                
+            time.sleep(interval)
+    
+    def _save_partial_data(self):
+        """Save partial data to prevent loss on crash"""
+        try:
+            partial_file = f"memory_profile_partial_{self.checkpoint_counter}.json"
+            with open(partial_file, 'w') as f:
+                json.dump({
+                    'timeline': self.data,
+                    'status': 'partial_save',
+                    'samples': len(self.data)
+                }, f, indent=2)
+            self.checkpoint_counter += 1
+        except Exception as e:
+            print(f"Failed to save partial data: {e}")
+    
+    def log_checkpoint(self, checkpoint_name: str):
+        """Log a specific checkpoint"""
+        if self.data:
+            self.data[-1]['checkpoint'] = checkpoint_name
+            latest = self.data[-1]
+            print(f"📍 CHECKPOINT [{checkpoint_name}]: RSS={latest['rss_gb']:.2f}GB, VRAM={latest['vram_gb']:.2f}GB")
+            
+            # Save checkpoint data immediately
+            self._save_partial_data()
+
+def run_with_profiling(config_path: str):
+    """Run the VR180 matting with memory profiling"""
+    profiler = MemoryProfiler("memory_profile_detailed.json")
+    
+    try:
+        # Start monitoring
+        profiler.start_monitoring(interval=2.0)  # Sample every 2 seconds
+        
+        # Log initial state
+        profiler.log_checkpoint("STARTUP")
+        
+        # Import after starting profiler to catch import memory usage
+        print("Importing VR180 processor...")
+        from vr180_matting.vr180_processor import VR180Processor
+        from vr180_matting.config import VR180Config
+        
+        profiler.log_checkpoint("IMPORTS_COMPLETE")
+        
+        # Load config
+        print(f"Loading config from {config_path}")
+        config = VR180Config.from_yaml(config_path)
+        
+        profiler.log_checkpoint("CONFIG_LOADED")
+        
+        # Initialize processor
+        print("Initializing VR180 processor...")
+        processor = VR180Processor(config)
+        
+        profiler.log_checkpoint("PROCESSOR_INITIALIZED")
+        
+        # Force garbage collection
+        gc.collect()
+        profiler.log_checkpoint("INITIAL_GC_COMPLETE")
+        
+        # Run processing
+        print("Starting VR180 processing...")
+        processor.process_video()
+        
+        profiler.log_checkpoint("PROCESSING_COMPLETE")
+        
+    except Exception as e:
+        print(f"❌ Error during processing: {e}")
+        profiler.log_checkpoint(f"ERROR: {str(e)}")
+        raise
+    finally:
+        # Stop monitoring and save results
+        profiler.stop_monitoring()
+        
+        # Print summary
+        print("\n" + "="*60)
+        print("MEMORY PROFILING SUMMARY")
+        print("="*60)
+        
+        if profiler.data:
+            peak_rss = max([d['rss_gb'] for d in profiler.data])
+            peak_vram = max([d['vram_gb'] for d in profiler.data])
+            
+            print(f"Peak RSS Memory: {peak_rss:.2f} GB")
+            print(f"Peak VRAM Usage: {peak_vram:.2f} GB")
+            print(f"Total Samples: {len(profiler.data)}")
+            
+            # Show checkpoints
+            checkpoints = [d for d in profiler.data if 'checkpoint' in d]
+            if checkpoints:
+                print(f"\nCheckpoints ({len(checkpoints)}):")
+                for cp in checkpoints:
+                    print(f"  {cp['checkpoint']}: RSS={cp['rss_gb']:.2f}GB, VRAM={cp['vram_gb']:.2f}GB")
+        
+        print(f"\nDetailed profile saved to: {profiler.output_file}")
+
+def main():
+    if len(sys.argv) != 2:
+        print("Usage: python memory_profiler_script.py <config.yaml>")
+        print("\nThis script runs VR180 matting with detailed memory profiling")
+        print("It will:")
+        print("- Monitor RSS, VRAM, and system memory every 2 seconds")
+        print("- Track memory allocations with tracemalloc")
+        print("- Log checkpoints at key processing stages")
+        print("- Save detailed JSON report for analysis")
+        sys.exit(1)
+    
+    config_path = sys.argv[1]
+    
+    if not Path(config_path).exists():
+        print(f"❌ Config file not found: {config_path}")
+        sys.exit(1)
+    
+    print("🚀 Starting VR180 Memory Profiling")
+    print(f"Config: {config_path}")
+    print("="*60)
+    
+    run_with_profiling(config_path)
+
+if __name__ == "__main__":
+    main()

quick_memory_check.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+"""
+Quick memory and system check before running full pipeline
+"""
+
+import psutil
+import subprocess
+import sys
+from pathlib import Path
+
+def check_system():
+    """Check system resources before starting"""
+    print("🔍 SYSTEM RESOURCE CHECK")
+    print("=" * 50)
+    
+    # Memory info
+    memory = psutil.virtual_memory()
+    print(f"📊 RAM:")
+    print(f"   Total: {memory.total / (1024**3):.1f} GB")
+    print(f"   Available: {memory.available / (1024**3):.1f} GB")
+    print(f"   Used: {(memory.total - memory.available) / (1024**3):.1f} GB ({memory.percent:.1f}%)")
+    
+    # GPU info
+    try:
+        result = subprocess.run(['nvidia-smi', '--query-gpu=name,memory.total,memory.used,memory.free', 
+                               '--format=csv,noheader,nounits'], 
+                              capture_output=True, text=True, timeout=10)
+        if result.returncode == 0:
+            lines = result.stdout.strip().split('\n')
+            print(f"\n🎮 GPU:")
+            for i, line in enumerate(lines):
+                if line.strip():
+                    parts = line.split(', ')
+                    if len(parts) >= 4:
+                        name, total, used, free = parts[:4]
+                        total_gb = float(total) / 1024
+                        used_gb = float(used) / 1024
+                        free_gb = float(free) / 1024
+                        print(f"   GPU {i}: {name}")
+                        print(f"   VRAM: {used_gb:.1f}/{total_gb:.1f} GB ({used_gb/total_gb*100:.1f}% used)")
+                        print(f"   Free: {free_gb:.1f} GB")
+    except Exception as e:
+        print(f"\n⚠️  Could not get GPU info: {e}")
+    
+    # Disk space
+    disk = psutil.disk_usage('/')
+    print(f"\n💾 Disk (/):")
+    print(f"   Total: {disk.total / (1024**3):.1f} GB")
+    print(f"   Used: {disk.used / (1024**3):.1f} GB ({disk.used/disk.total*100:.1f}%)")
+    print(f"   Free: {disk.free / (1024**3):.1f} GB")
+    
+    # Check config file
+    if len(sys.argv) > 1:
+        config_path = sys.argv[1]
+        if Path(config_path).exists():
+            print(f"\n✅ Config file found: {config_path}")
+            
+            # Try to load and show key settings
+            try:
+                import yaml
+                with open(config_path, 'r') as f:
+                    config = yaml.safe_load(f)
+                
+                print(f"📋 Key Settings:")
+                if 'processing' in config:
+                    proc = config['processing']
+                    print(f"   Chunk size: {proc.get('chunk_size', 'default')}")
+                    print(f"   Scale factor: {proc.get('scale_factor', 'default')}")
+                
+                if 'hardware' in config:
+                    hw = config['hardware']
+                    print(f"   Max VRAM: {hw.get('max_vram_gb', 'default')} GB")
+                    
+                if 'input' in config:
+                    inp = config['input']
+                    video_path = inp.get('video_path', '')
+                    if video_path and Path(video_path).exists():
+                        size_gb = Path(video_path).stat().st_size / (1024**3)
+                        print(f"   Input video: {video_path} ({size_gb:.1f} GB)")
+                    else:
+                        print(f"   ⚠️  Input video not found: {video_path}")
+                        
+            except Exception as e:
+                print(f"   ⚠️  Could not parse config: {e}")
+        else:
+            print(f"\n❌ Config file not found: {config_path}")
+            return False
+    
+    # Memory safety warnings
+    print(f"\n⚠️  MEMORY SAFETY CHECKS:")
+    available_gb = memory.available / (1024**3)
+    
+    if available_gb < 10:
+        print(f"   🔴 LOW MEMORY: Only {available_gb:.1f}GB available")
+        print("      Consider: reducing chunk_size or scale_factor")
+        return False
+    elif available_gb < 20:
+        print(f"   🟡 MODERATE MEMORY: {available_gb:.1f}GB available")
+        print("      Recommend: chunk_size ≤ 300, scale_factor ≤ 0.5")
+    else:
+        print(f"   🟢 GOOD MEMORY: {available_gb:.1f}GB available")
+    
+    print(f"\n" + "=" * 50)
+    return True
+
+def main():
+    if len(sys.argv) != 2:
+        print("Usage: python quick_memory_check.py <config.yaml>")
+        print("\nThis checks system resources before running VR180 matting")
+        sys.exit(1)
+    
+    safe_to_run = check_system()
+    
+    if safe_to_run:
+        print("✅ System check passed - safe to run VR180 matting")
+        print("\nTo run with memory profiling:")
+        print(f"   python memory_profiler_script.py {sys.argv[1]}")
+        print("\nTo run normally:")
+        print(f"   vr180-matting {sys.argv[1]}")
+    else:
+        print("❌ System check failed - address issues before running")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -10,3 +10,6 @@ tqdm>=4.65.0
 psutil>=5.9.0
 ffmpeg-python>=0.2.0
 decord>=0.6.0
+# GPU acceleration (optional but recommended for stereo validation speedup)
+# cupy-cuda11x>=12.0.0  # For CUDA 11.x
+# cupy-cuda12x>=12.0.0  # For CUDA 12.x - uncomment appropriate version

runpod_setup.sh

@@ -18,6 +18,28 @@ pip install -r requirements.txt
 echo "📹 Installing decord for video processing..."
 pip install decord
 
+# Install CuPy for GPU acceleration of stereo validation
+echo "🚀 Installing CuPy for GPU acceleration..."
+# Auto-detect CUDA version and install appropriate CuPy
+python -c "
+import torch
+if torch.cuda.is_available():
+    cuda_version = torch.version.cuda
+    print(f'CUDA version detected: {cuda_version}')
+    if cuda_version.startswith('11.'):
+        import subprocess
+        subprocess.run(['pip', 'install', 'cupy-cuda11x>=12.0.0'])
+        print('Installed CuPy for CUDA 11.x')
+    elif cuda_version.startswith('12.'):
+        import subprocess
+        subprocess.run(['pip', 'install', 'cupy-cuda12x>=12.0.0'])
+        print('Installed CuPy for CUDA 12.x')
+    else:
+        print(f'Unsupported CUDA version: {cuda_version}')
+else:
+    print('CUDA not available, skipping CuPy installation')
+"
+
 # Install SAM2 separately (not on PyPI)
 echo "🎯 Installing SAM2..."
 pip install git+https://github.com/facebookresearch/segment-anything-2.git

test_inter_chunk_cleanup.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+"""
+Test script to verify inter-chunk cleanup properly destroys models
+"""
+
+import psutil
+import gc
+import sys
+from pathlib import Path
+
+def get_memory_usage():
+    """Get current memory usage in GB"""
+    process = psutil.Process()
+    return process.memory_info().rss / (1024**3)
+
+def test_inter_chunk_cleanup():
+    """Test that models are properly destroyed between chunks"""
+    
+    print("🧪 TESTING INTER-CHUNK CLEANUP")
+    print("=" * 50)
+    
+    baseline_memory = get_memory_usage()
+    print(f"📊 Baseline memory: {baseline_memory:.2f} GB")
+    
+    # Import and create processor
+    print("\n1️⃣ Creating processor...")
+    from vr180_matting.config import VR180Config
+    from vr180_matting.vr180_processor import VR180Processor
+    
+    config = VR180Config.from_yaml('config.yaml')
+    processor = VR180Processor(config)
+    
+    init_memory = get_memory_usage()
+    print(f"📊 After processor init: {init_memory:.2f} GB (+{init_memory - baseline_memory:.2f} GB)")
+    
+    # Simulate chunk processing (just trigger model loading)
+    print("\n2️⃣ Simulating chunk 0 processing...")
+    
+    # Test 1: Force YOLO model loading
+    try:
+        detector = processor.detector
+        detector._load_model()  # Force load
+        yolo_memory = get_memory_usage()
+        print(f"📊 After YOLO load: {yolo_memory:.2f} GB (+{yolo_memory - init_memory:.2f} GB)")
+    except Exception as e:
+        print(f"❌ YOLO loading failed: {e}")
+        yolo_memory = init_memory
+    
+    # Test 2: Force SAM2 model loading
+    try:
+        sam2_model = processor.sam2_model
+        sam2_model._load_model(sam2_model.model_cfg, sam2_model.checkpoint_path)
+        sam2_memory = get_memory_usage()
+        print(f"📊 After SAM2 load: {sam2_memory:.2f} GB (+{sam2_memory - yolo_memory:.2f} GB)")
+    except Exception as e:
+        print(f"❌ SAM2 loading failed: {e}")
+        sam2_memory = yolo_memory
+    
+    total_model_memory = sam2_memory - init_memory
+    print(f"📊 Total model memory: {total_model_memory:.2f} GB")
+    
+    # Test 3: Inter-chunk cleanup
+    print("\n3️⃣ Testing inter-chunk cleanup...")
+    processor._complete_inter_chunk_cleanup(chunk_idx=0)
+    
+    cleanup_memory = get_memory_usage()
+    cleanup_improvement = sam2_memory - cleanup_memory
+    print(f"📊 After cleanup: {cleanup_memory:.2f} GB (-{cleanup_improvement:.2f} GB freed)")
+    
+    # Test 4: Verify models reload fresh
+    print("\n4️⃣ Testing fresh model reload...")
+    
+    # Check YOLO state
+    yolo_reloaded = processor.detector.model is None
+    print(f"🔍 YOLO model destroyed: {'✅ YES' if yolo_reloaded else '❌ NO'}")
+    
+    # Check SAM2 state  
+    sam2_reloaded = not processor.sam2_model._model_loaded or processor.sam2_model.predictor is None
+    print(f"🔍 SAM2 model destroyed: {'✅ YES' if sam2_reloaded else '❌ NO'}")
+    
+    # Test 5: Force reload to verify they work
+    print("\n5️⃣ Testing model reload...")
+    try:
+        # Force YOLO reload
+        processor.detector._load_model()
+        yolo_reload_memory = get_memory_usage()
+        
+        # Force SAM2 reload
+        processor.sam2_model._load_model(processor.sam2_model.model_cfg, processor.sam2_model.checkpoint_path)
+        sam2_reload_memory = get_memory_usage()
+        
+        reload_growth = sam2_reload_memory - cleanup_memory
+        print(f"📊 After reload: {sam2_reload_memory:.2f} GB (+{reload_growth:.2f} GB)")
+        
+        if abs(reload_growth - total_model_memory) < 1.0:  # Within 1GB
+            print("✅ Models reloaded with similar memory usage (good)")
+        else:
+            print("⚠️  Model reload memory differs significantly")
+            
+    except Exception as e:
+        print(f"❌ Model reload failed: {e}")
+    
+    # Final summary
+    print(f"\n📊 SUMMARY:")
+    print(f"   Baseline → Peak: {baseline_memory:.2f}GB → {sam2_memory:.2f}GB")
+    print(f"   Peak → Cleanup: {sam2_memory:.2f}GB → {cleanup_memory:.2f}GB")
+    print(f"   Memory freed: {cleanup_improvement:.2f}GB")
+    print(f"   Models destroyed: YOLO={yolo_reloaded}, SAM2={sam2_reloaded}")
+    
+    if cleanup_improvement > total_model_memory * 0.5:  # Freed >50% of model memory
+        print("✅ Inter-chunk cleanup working effectively")
+        return True
+    else:
+        print("❌ Inter-chunk cleanup not freeing enough memory")
+        return False
+
+def main():
+    if len(sys.argv) != 2:
+        print("Usage: python test_inter_chunk_cleanup.py <config.yaml>")
+        sys.exit(1)
+    
+    config_path = sys.argv[1]
+    if not Path(config_path).exists():
+        print(f"Config file not found: {config_path}")
+        sys.exit(1)
+    
+    success = test_inter_chunk_cleanup()
+    
+    if success:
+        print(f"\n🎉 SUCCESS: Inter-chunk cleanup is working!")
+        print(f"💡 This should prevent 15-20GB model accumulation between chunks")
+    else:
+        print(f"\n❌ FAILURE: Inter-chunk cleanup needs improvement")
+        print(f"💡 Check model destruction logic in _complete_inter_chunk_cleanup")
+    
+    return 0 if success else 1
+
+if __name__ == "__main__":
+    sys.exit(main())

vr180_matting/config.py

@@ -29,6 +29,11 @@ class MattingConfig:
     fp16: bool = True
     sam2_model_cfg: str = "sam2.1_hiera_l"
     sam2_checkpoint: str = "segment-anything-2/checkpoints/sam2.1_hiera_large.pt"
+    # Det-SAM2 optimizations
+    continuous_correction: bool = True
+    correction_interval: int = 60  # Add correction prompts every N frames
+    frame_release_interval: int = 50  # Release old frames every N frames
+    frame_window_size: int = 30  # Keep N frames in memory
 
 
 @dataclass

vr180_matting/detector.py

@@ -1,6 +1,4 @@
-import torch
 import numpy as np
-from ultralytics import YOLO
 from typing import List, Tuple, Dict, Any
 import cv2
 
@@ -13,14 +11,23 @@ class YOLODetector:
         self.confidence_threshold = confidence_threshold
         self.device = device
         self.model = None
-        self._load_model()
+        # Don't load model during init - load lazily when first used
     
     def _load_model(self):
-        """Load YOLOv8 model"""
+        """Load YOLOv8 model lazily"""
+        if self.model is not None:
+            return  # Already loaded
+            
         try:
+            # Import heavy dependencies only when needed
+            import torch
+            from ultralytics import YOLO
+            
             self.model = YOLO(f"{self.model_name}.pt")
             if self.device == "cuda" and torch.cuda.is_available():
                 self.model.to("cuda")
+                
+            print(f"🎯 Loaded YOLO model: {self.model_name}")
         except Exception as e:
             raise RuntimeError(f"Failed to load YOLO model {self.model_name}: {e}")
     
@@ -34,8 +41,9 @@ class YOLODetector:
         Returns:
             List of detection dictionaries with bbox, confidence, and class info
         """
+        # Load model lazily on first use
         if self.model is None:
-            raise RuntimeError("YOLO model not loaded")
+            self._load_model()
         
         results = self.model(frame, verbose=False)
         detections = []

vr180_matting/sam2_wrapper.py

@@ -7,13 +7,18 @@ import warnings
 import os
 import tempfile
 import shutil
+import gc
 
-try:
-    from sam2.build_sam import build_sam2_video_predictor
-    from sam2.sam2_image_predictor import SAM2ImagePredictor
-    SAM2_AVAILABLE = True
-except ImportError:
-    SAM2_AVAILABLE = False
+# Check SAM2 availability without importing heavy modules
+def _check_sam2_available():
+    try:
+        import sam2
+        return True
+    except ImportError:
+        return False
+
+SAM2_AVAILABLE = _check_sam2_available()
+if not SAM2_AVAILABLE:
     warnings.warn("SAM2 not available. Please install sam2 package.")
 
 
@@ -32,16 +37,25 @@ class SAM2VideoMatting:
         self.device = device
         self.memory_offload = memory_offload
         self.fp16 = fp16
+        self.model_cfg = model_cfg
+        self.checkpoint_path = checkpoint_path
         self.predictor = None
         self.inference_state = None
         self.video_segments = {}
         self.temp_video_path = None
         
-        self._load_model(model_cfg, checkpoint_path)
+        # Don't load model during init - load lazily when needed
+        self._model_loaded = False
     
     def _load_model(self, model_cfg: str, checkpoint_path: str):
-        """Load SAM2 video predictor with optimizations"""
+        """Load SAM2 video predictor lazily"""
+        if self._model_loaded and self.predictor is not None:
+            return  # Already loaded and predictor exists
+            
         try:
+            # Import heavy SAM2 modules only when needed
+            from sam2.build_sam import build_sam2_video_predictor
+            
             # Check for checkpoint in SAM2 repo structure
             if not Path(checkpoint_path).exists():
                 # Try in segment-anything-2/checkpoints/
@@ -60,6 +74,7 @@ class SAM2VideoMatting:
                         if sam2_repo_path.exists():
                             checkpoint_path = str(sam2_repo_path)
             
+            print(f"🎯 Loading SAM2 model: {model_cfg}")
             # Use SAM2's build_sam2_video_predictor which returns the predictor directly
             # The predictor IS the model - no .model attribute needed
             self.predictor = build_sam2_video_predictor(
@@ -68,13 +83,17 @@ class SAM2VideoMatting:
                 device=self.device
             )
             
+            self._model_loaded = True
+            print(f"✅ SAM2 model loaded successfully")
+                
         except Exception as e:
             raise RuntimeError(f"Failed to load SAM2 model: {e}")
     
     def init_video_state(self, video_frames: List[np.ndarray] = None, video_path: str = None) -> None:
         """Initialize video inference state"""
-        if self.predictor is None:
-            raise RuntimeError("SAM2 model not loaded")
+        # Load model lazily on first use
+        if not self._model_loaded:
+            self._load_model(self.model_cfg, self.checkpoint_path)
         
         if video_path is not None:
             # Use video path directly (SAM2's preferred method)
@@ -148,13 +167,16 @@ class SAM2VideoMatting:
         
         return object_ids
     
-    def propagate_masks(self, start_frame: int = 0, max_frames: Optional[int] = None) -> Dict[int, Dict[int, np.ndarray]]:
+    def propagate_masks(self, start_frame: int = 0, max_frames: Optional[int] = None, 
+                       frame_release_interval: int = 50, frame_window_size: int = 30) -> Dict[int, Dict[int, np.ndarray]]:
         """
-        Propagate masks through video
+        Propagate masks through video with Det-SAM2 style memory management
         
         Args:
             start_frame: Starting frame index
             max_frames: Maximum number of frames to process
+            frame_release_interval: Release old frames every N frames
+            frame_window_size: Keep N frames in memory
             
         Returns:
             Dictionary mapping frame_idx -> {obj_id: mask}
@@ -178,9 +200,108 @@ class SAM2VideoMatting:
             
             video_segments[out_frame_idx] = frame_masks
             
-            # Memory management: release old frames periodically
-            if self.memory_offload and out_frame_idx % 100 == 0:
-                self._release_old_frames(out_frame_idx - 50)
+            # Det-SAM2 style memory management: more aggressive frame release
+            if self.memory_offload and out_frame_idx % frame_release_interval == 0:
+                self._release_old_frames(out_frame_idx - frame_window_size)
+                # Optional: Log frame release for monitoring
+                if out_frame_idx % (frame_release_interval * 4) == 0:  # Log every 4x release interval
+                    print(f"Det-SAM2: Released frames before {out_frame_idx - frame_window_size}, keeping {frame_window_size} frames")
+        
+        return video_segments
+    
+    def propagate_masks_with_continuous_correction(self, 
+                                                 detector,
+                                                 temp_video_path: str,
+                                                 start_frame: int = 0, 
+                                                 max_frames: Optional[int] = None,
+                                                 correction_interval: int = 60,
+                                                 frame_release_interval: int = 50,
+                                                 frame_window_size: int = 30) -> Dict[int, Dict[int, np.ndarray]]:
+        """
+        Det-SAM2 style: Propagate masks with continuous prompt correction
+        
+        Args:
+            detector: YOLODetector instance for generating correction prompts
+            temp_video_path: Path to video file for frame access
+            start_frame: Starting frame index
+            max_frames: Maximum number of frames to process
+            correction_interval: Add correction prompts every N frames
+            frame_release_interval: Release old frames every N frames
+            frame_window_size: Keep N frames in memory
+            
+        Returns:
+            Dictionary mapping frame_idx -> {obj_id: mask}
+        """
+        if self.inference_state is None:
+            raise RuntimeError("Video state not initialized")
+        
+        video_segments = {}
+        max_frames = max_frames or 10000  # Default limit
+        
+        # Open video for accessing frames during propagation
+        cap = cv2.VideoCapture(str(temp_video_path))
+        
+        try:
+            for out_frame_idx, out_obj_ids, out_mask_logits in self.predictor.propagate_in_video(
+                self.inference_state, 
+                start_frame_idx=start_frame,
+                max_frame_num_to_track=max_frames,
+                reverse=False
+            ):
+                frame_masks = {}
+                
+                for i, out_obj_id in enumerate(out_obj_ids):
+                    mask = (out_mask_logits[i] > 0.0).cpu().numpy()
+                    frame_masks[out_obj_id] = mask
+                
+                video_segments[out_frame_idx] = frame_masks
+                
+                # Det-SAM2 optimization: Add correction prompts at keyframes
+                if (out_frame_idx % correction_interval == 0 and 
+                    out_frame_idx > start_frame and 
+                    out_frame_idx < max_frames - 1):
+                    
+                    # Read frame for detection
+                    cap.set(cv2.CAP_PROP_POS_FRAMES, out_frame_idx)
+                    ret, correction_frame = cap.read()
+                    
+                    if ret:
+                        # Run detection on this keyframe
+                        detections = detector.detect_persons(correction_frame)
+                        
+                        if detections:
+                            # Convert to prompts and add as corrections
+                            box_prompts, labels = detector.convert_to_sam_prompts(detections)
+                            
+                            # Add correction prompts (SAM2 will propagate backward)
+                            correction_count = 0
+                            try:
+                                for i, (box, label) in enumerate(zip(box_prompts, labels)):
+                                    # Use existing object IDs if available, otherwise create new ones
+                                    obj_id = out_obj_ids[i] if i < len(out_obj_ids) else len(out_obj_ids) + i + 1
+                                    
+                                    self.predictor.add_new_points_or_box(
+                                        inference_state=self.inference_state,
+                                        frame_idx=out_frame_idx,
+                                        obj_id=obj_id,
+                                        box=box,
+                                    )
+                                    correction_count += 1
+                                
+                                print(f"Det-SAM2: Added {correction_count} correction prompts at frame {out_frame_idx}")
+                                
+                            except Exception as e:
+                                warnings.warn(f"Failed to add correction prompt at frame {out_frame_idx}: {e}")
+                
+                # Memory management: More aggressive frame release (Det-SAM2 style)
+                if self.memory_offload and out_frame_idx % frame_release_interval == 0:
+                    self._release_old_frames(out_frame_idx - frame_window_size)
+                    # Optional: Log frame release for monitoring
+                    if out_frame_idx % (frame_release_interval * 4) == 0:  # Log every 4x release interval
+                        print(f"Det-SAM2: Released frames before {out_frame_idx - frame_window_size}, keeping {frame_window_size} frames")
+            
+        finally:
+            cap.release()
         
         return video_segments
     
@@ -256,11 +377,23 @@ class SAM2VideoMatting:
         """Clean up resources"""
         if self.inference_state is not None:
             try:
-                if hasattr(self.predictor, 'cleanup_state'):
+                # Reset SAM2 state first (critical for memory cleanup)
+                if self.predictor is not None and hasattr(self.predictor, 'reset_state'):
+                    self.predictor.reset_state(self.inference_state)
+                
+                # Fallback to cleanup_state if available
+                elif self.predictor is not None and hasattr(self.predictor, 'cleanup_state'):
                     self.predictor.cleanup_state(self.inference_state)
+                
+                # Explicitly delete inference state and video segments
+                del self.inference_state
+                if hasattr(self, 'video_segments') and self.video_segments:
+                    del self.video_segments
+                    self.video_segments = {}
+                
             except Exception as e:
                 warnings.warn(f"Failed to cleanup SAM2 state: {e}")
-            
+            finally:
                 self.inference_state = None
         
         # Clean up temporary video file
@@ -277,6 +410,25 @@ class SAM2VideoMatting:
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
         
+        # Explicitly delete predictor for fresh creation next time
+        if self.predictor is not None:
+            try:
+                del self.predictor
+            except Exception as e:
+                warnings.warn(f"Failed to delete predictor: {e}")
+            finally:
+                self.predictor = None
+        
+        # Reset model loaded state for fresh reload
+        self._model_loaded = False
+        
+        # Force garbage collection (critical for memory leak prevention)
+        gc.collect()
+    
     def __del__(self):
         """Destructor to ensure cleanup"""
+        try:
             self.cleanup()
+        except Exception:
+            # Ignore errors during Python shutdown
+            pass

vr180_matting/video_processor.py

@@ -132,6 +132,26 @@ class VideoProcessor:
         except ImportError:
             pass
         
+        # Clear OpenCV internal caches
+        try:
+            # Clear OpenCV video capture cache
+            cv2.setUseOptimized(False)
+            cv2.setUseOptimized(True)
+        except Exception:
+            pass
+        
+        # Clear CuPy caches if available  
+        try:
+            import cupy as cp
+            cp._default_memory_pool.free_all_blocks()
+            cp._default_pinned_memory_pool.free_all_blocks()
+            cp.get_default_memory_pool().free_all_blocks()
+            cp.get_default_pinned_memory_pool().free_all_blocks()
+        except ImportError:
+            pass
+        except Exception as e:
+            print(f"   Warning: Could not clear CuPy cache: {e}")
+        
         # Force Linux to release memory back to OS
         if sys.platform == 'linux':
             try:
@@ -367,19 +387,83 @@ class VideoProcessor:
             # Green screen background
             return np.full_like(frame, self.config.output.background_color, dtype=np.uint8)
     
+    def merge_chunks_streaming(self, chunk_files: List[Path], output_path: str, 
+                             overlap_frames: int = 0, audio_source: str = None) -> None:
+        """
+        Merge processed chunks using streaming approach (no memory accumulation)
+        
+        Args:
+            chunk_files: List of chunk result files (.npz)
+            output_path: Final output video path
+            overlap_frames: Number of overlapping frames
+            audio_source: Audio source file for final video
+        """
+        from .streaming_video_writer import StreamingVideoWriter
+        
+        if not chunk_files:
+            raise ValueError("No chunk files to merge")
+            
+        print(f"🎬 Streaming merge: {len(chunk_files)} chunks → {output_path}")
+        
+        # Initialize streaming writer
+        writer = StreamingVideoWriter(
+            output_path=output_path,
+            fps=self.video_info['fps'],
+            audio_source=audio_source
+        )
+        
+        try:
+            # Process each chunk without accumulation
+            for i, chunk_file in enumerate(chunk_files):
+                print(f"📼 Processing chunk {i+1}/{len(chunk_files)}: {chunk_file.name}")
+                
+                # Load chunk (this is the only copy in memory)
+                chunk_data = np.load(str(chunk_file))
+                frames = chunk_data['frames'].tolist()  # Convert to list of arrays
+                chunk_data.close()
+                
+                # Write chunk with streaming writer
+                writer.write_chunk(
+                    frames=frames,
+                    chunk_index=i,
+                    overlap_frames=overlap_frames if i > 0 else 0,
+                    blend_with_previous=(i > 0 and overlap_frames > 0)
+                )
+                
+                # Immediately free memory
+                del frames, chunk_data
+                
+                # Delete chunk file to free disk space
+                try:
+                    chunk_file.unlink()
+                    print(f"   🗑️  Deleted {chunk_file.name}")
+                except Exception as e:
+                    print(f"   ⚠️  Could not delete {chunk_file.name}: {e}")
+                
+                # Aggressive cleanup every chunk
+                self._aggressive_memory_cleanup(f"After processing chunk {i}")
+                
+            # Finalize the video
+            writer.finalize()
+            
+        except Exception as e:
+            print(f"❌ Streaming merge failed: {e}")
+            writer.cleanup()
+            raise
+            
+        print(f"✅ Streaming merge complete: {output_path}")
+    
     def merge_overlapping_chunks(self, 
                                 chunk_results: List[List[np.ndarray]], 
                                 overlap_frames: int) -> List[np.ndarray]:
         """
-        Merge overlapping chunks with blending in overlap regions
-        
-        Args:
-            chunk_results: List of chunk results
-            overlap_frames: Number of overlapping frames
-            
-        Returns:
-            Merged frame sequence
+        Legacy merge method - DEPRECATED due to memory accumulation
+        Use merge_chunks_streaming() instead for memory efficiency
         """
+        import warnings
+        warnings.warn("merge_overlapping_chunks() is deprecated due to memory accumulation. Use merge_chunks_streaming()", 
+                     DeprecationWarning, stacklevel=2)
+        
         if len(chunk_results) == 1:
             return chunk_results[0]
         
@@ -620,25 +704,23 @@ class VideoProcessor:
                 if self.memory_manager.should_emergency_cleanup():
                     self.memory_manager.emergency_cleanup()
             
-            # Load and merge chunks from disk
-            print("\nLoading and merging chunks...")
-            chunk_results = []
-            for chunk_file in chunk_files:
-                print(f"Loading {chunk_file.name}...")
-                chunk_data = np.load(str(chunk_file))
-                chunk_results.append(chunk_data['frames'])
-                chunk_data.close()  # Close the file
+            # Use streaming merge to avoid memory accumulation (fixes OOM)
+            print("\n🎬 Using streaming merge (no memory accumulation)...")
             
-            # Merge chunks
-            final_frames = self.merge_overlapping_chunks(chunk_results, overlap_frames)
+            # Determine audio source for final video
+            audio_source = None
+            if self.config.output.preserve_audio and Path(self.config.input.video_path).exists():
+                audio_source = self.config.input.video_path
             
-            # Free chunk results after merging
-            del chunk_results
-            self._aggressive_memory_cleanup("after merging chunks")
+            # Stream merge chunks directly to output (no memory accumulation)
+            self.merge_chunks_streaming(
+                chunk_files=chunk_files,
+                output_path=self.config.output.path,
+                overlap_frames=overlap_frames,
+                audio_source=audio_source
+            )
             
-            # Save results
-            print(f"Saving {len(final_frames)} processed frames...")
-            self.save_video(final_frames, self.config.output.path)
+            print("✅ Streaming merge complete - no memory accumulation!")
             
             # Calculate final statistics
             self.processing_stats['end_time'] = time.time()

vr180_matting/vr180_processor.py

@@ -3,6 +3,7 @@ import numpy as np
 from typing import List, Dict, Any, Optional, Tuple
 from pathlib import Path
 import warnings
+import torch
 
 from .video_processor import VideoProcessor
 from .config import VR180Config
@@ -89,7 +90,7 @@ class VR180Processor(VideoProcessor):
     
     def combine_sbs_frame(self, left_eye: np.ndarray, right_eye: np.ndarray) -> np.ndarray:
         """
-        Combine left and right eye frames back into side-by-side format
+        Combine left and right eye frames back into side-by-side format with GPU acceleration
         
         Args:
             left_eye: Left eye frame
@@ -98,14 +99,44 @@ class VR180Processor(VideoProcessor):
         Returns:
             Combined SBS frame
         """
+        try:
+            import cupy as cp
+            
+            # Transfer to GPU for faster combination
+            left_gpu = cp.asarray(left_eye)
+            right_gpu = cp.asarray(right_eye)
+            
+            # Ensure frames have same height
+            if left_gpu.shape[0] != right_gpu.shape[0]:
+                target_height = min(left_gpu.shape[0], right_gpu.shape[0])
+                # Note: OpenCV resize not available in CuPy, fall back to CPU for resize
+                left_eye = cv2.resize(left_eye, (left_eye.shape[1], target_height))
+                right_eye = cv2.resize(right_eye, (right_eye.shape[1], target_height))
+                left_gpu = cp.asarray(left_eye)
+                right_gpu = cp.asarray(right_eye)
+            
+            # Combine horizontally on GPU (much faster for large arrays)
+            combined_gpu = cp.hstack([left_gpu, right_gpu])
+            
+            # Transfer back to CPU and ensure we get a copy, not a view
+            combined = cp.asnumpy(combined_gpu).copy()
+            
+            # Free GPU memory immediately
+            del left_gpu, right_gpu, combined_gpu
+            cp._default_memory_pool.free_all_blocks()
+            
+            return combined
+            
+        except ImportError:
+            # Fallback to CPU NumPy
             # Ensure frames have same height
             if left_eye.shape[0] != right_eye.shape[0]:
                 target_height = min(left_eye.shape[0], right_eye.shape[0])
                 left_eye = cv2.resize(left_eye, (left_eye.shape[1], target_height))
                 right_eye = cv2.resize(right_eye, (right_eye.shape[1], target_height))
             
-        # Combine horizontally
-        combined = np.hstack([left_eye, right_eye])
+            # Combine horizontally and ensure we get a copy, not a view
+            combined = np.hstack([left_eye, right_eye]).copy()
             return combined
     
     def process_with_disparity_mapping(self, 
@@ -152,6 +183,10 @@ class VR180Processor(VideoProcessor):
         with self.memory_manager.memory_monitor(f"left eye chunk {chunk_idx}"):
             left_matted = self._process_eye_sequence(left_eye_frames, "left", chunk_idx)
         
+        # Free left eye frames after processing (before right eye to save memory)
+        del left_eye_frames
+        self._aggressive_memory_cleanup(f"After left eye processing chunk {chunk_idx}")
+        
         # Process right eye with cross-validation
         print("Processing right eye with cross-validation...")
         with self.memory_manager.memory_monitor(f"right eye chunk {chunk_idx}"):
@@ -159,6 +194,10 @@ class VR180Processor(VideoProcessor):
                 right_eye_frames, left_matted, "right", chunk_idx
             )
         
+        # Free right eye frames after processing
+        del right_eye_frames
+        self._aggressive_memory_cleanup(f"After right eye processing chunk {chunk_idx}")
+        
         # Combine results back to SBS format
         combined_frames = []
         for left_frame, right_frame in zip(left_matted, right_matted):
@@ -169,6 +208,15 @@ class VR180Processor(VideoProcessor):
                 combined = {'left': left_frame, 'right': right_frame}
             combined_frames.append(combined)
         
+        # Free the individual eye results after combining
+        del left_matted
+        del right_matted
+        self._aggressive_memory_cleanup(f"After combining frames chunk {chunk_idx}")
+        
+        # CRITICAL: Complete inter-chunk cleanup to prevent model persistence
+        # This ensures models don't accumulate between chunks
+        self._complete_inter_chunk_cleanup(chunk_idx)
+        
         return combined_frames
     
     def _process_eye_sequence(self, 
@@ -332,31 +380,43 @@ class VR180Processor(VideoProcessor):
             
             # Propagate masks (most expensive operation)
             self._print_memory_step(f"Before SAM2 propagation ({eye_name} eye, {num_frames} frames)")
+            
+            # Use Det-SAM2 continuous correction if enabled
+            if self.config.matting.continuous_correction:
+                video_segments = self.sam2_model.propagate_masks_with_continuous_correction(
+                    detector=self.detector,
+                    temp_video_path=str(temp_video_path),
+                    start_frame=0, 
+                    max_frames=num_frames,
+                    correction_interval=self.config.matting.correction_interval,
+                    frame_release_interval=self.config.matting.frame_release_interval,
+                    frame_window_size=self.config.matting.frame_window_size
+                )
+                print(f"Used Det-SAM2 continuous correction (interval: {self.config.matting.correction_interval} frames)")
+            else:
                 video_segments = self.sam2_model.propagate_masks(
                     start_frame=0, 
-                max_frames=num_frames
+                    max_frames=num_frames,
+                    frame_release_interval=self.config.matting.frame_release_interval,
+                    frame_window_size=self.config.matting.frame_window_size
                 )
+            
             self._print_memory_step(f"After SAM2 propagation ({eye_name} eye)")
             
-            # Apply masks - need to reload frames from temp video since we freed the original frames
-            self._print_memory_step(f"Before reloading frames for mask application ({eye_name} eye)")
+            # Apply masks with streaming approach (no frame accumulation)
+            self._print_memory_step(f"Before streaming mask application ({eye_name} eye)")
             
-            # Read frames back from the temp video for mask application
+            # Process frames one at a time without accumulation
             cap = cv2.VideoCapture(str(temp_video_path))
-            reloaded_frames = []
+            matted_frames = []
             
+            try:
                 for frame_idx in range(num_frames):
                     ret, frame = cap.read()
                     if not ret:
                         break
-                reloaded_frames.append(frame)
-            cap.release()
                     
-            self._print_memory_step(f"Reloaded {len(reloaded_frames)} frames for mask application")
-            
-            # Apply masks
-            matted_frames = []
-            for frame_idx, frame in enumerate(reloaded_frames):
+                    # Apply mask to this single frame
                     if frame_idx in video_segments:
                         frame_masks = video_segments[frame_idx]
                         combined_mask = self.sam2_model.get_combined_mask(frame_masks)
@@ -371,10 +431,22 @@ class VR180Processor(VideoProcessor):
                     
                     matted_frames.append(matted_frame)
                     
-            # Free reloaded frames 
-            del reloaded_frames
-            self._aggressive_memory_cleanup(f"After mask application ({eye_name} eye)")
+                    # Free the original frame immediately (no accumulation)
+                    del frame
                     
+                    # Periodic cleanup during processing
+                    if frame_idx % 100 == 0 and frame_idx > 0:
+                        import gc
+                        gc.collect()
+                        
+            finally:
+                cap.release()
+            
+            # Free video segments completely
+            del video_segments  # This holds processed masks from SAM2
+            self._aggressive_memory_cleanup(f"After streaming mask application ({eye_name} eye)")
+            
+            self._print_memory_step(f"Completed streaming mask application ({eye_name} eye)")
             return matted_frames
             
         finally:
@@ -414,13 +486,17 @@ class VR180Processor(VideoProcessor):
             left_eye_results, right_matted
         )
         
+        # CRITICAL: Free the intermediate results to prevent memory accumulation
+        del left_eye_results  # Don't keep left eye results after validation
+        del right_matted     # Don't keep unvalidated right results
+        
         return validated_results
     
     def _validate_stereo_consistency(self, 
                                    left_results: List[np.ndarray], 
                                    right_results: List[np.ndarray]) -> List[np.ndarray]:
         """
-        Validate and correct stereo consistency between left and right eye results
+        Validate and correct stereo consistency between left and right eye results using GPU acceleration
         
         Args:
             left_results: Left eye processed frames
@@ -429,9 +505,120 @@ class VR180Processor(VideoProcessor):
         Returns:
             Validated right eye frames
         """
+        print(f"🔍 VALIDATION: Starting stereo consistency check ({len(left_results)} frames)")
+        
+        try:
+            import cupy as cp
+            return self._validate_stereo_consistency_gpu(left_results, right_results)
+        except ImportError:
+            print("   Warning: CuPy not available, using CPU validation")
+            return self._validate_stereo_consistency_cpu(left_results, right_results)
+    
+    def _validate_stereo_consistency_gpu(self, 
+                                       left_results: List[np.ndarray], 
+                                       right_results: List[np.ndarray]) -> List[np.ndarray]:
+        """GPU-accelerated batch stereo validation using CuPy with memory-safe batching"""
+        import cupy as cp
+        
+        print("   Using GPU acceleration for stereo validation")
+        
+        # Process in batches to avoid GPU OOM
+        batch_size = 50  # Process 50 frames at a time (safe for 45GB GPU)
+        total_frames = len(left_results)
+        area_ratios_all = []
+        needs_correction_all = []
+        
+        print(f"   Processing {total_frames} frames in batches of {batch_size}...")
+        
+        for batch_start in range(0, total_frames, batch_size):
+            batch_end = min(batch_start + batch_size, total_frames)
+            batch_frames = batch_end - batch_start
+            
+            if batch_start % 100 == 0:
+                print(f"   GPU batch {batch_start//batch_size + 1}: frames {batch_start}-{batch_end}")
+            
+            # Get batch slices
+            left_batch = left_results[batch_start:batch_end]
+            right_batch = right_results[batch_start:batch_end]
+            
+            # Convert batch to GPU
+            left_stack = cp.stack([cp.asarray(frame) for frame in left_batch])
+            right_stack = cp.stack([cp.asarray(frame) for frame in right_batch])
+            
+            # Batch calculate mask areas for this batch
+            if left_stack.shape[3] == 4:  # Alpha channel
+                left_masks = left_stack[:, :, :, 3] > 0
+                right_masks = right_stack[:, :, :, 3] > 0
+            else:  # Green screen detection
+                bg_color = cp.array(self.config.output.background_color)
+                left_diff = cp.abs(left_stack.astype(cp.float32) - bg_color).sum(axis=3)
+                right_diff = cp.abs(right_stack.astype(cp.float32) - bg_color).sum(axis=3)
+                left_masks = left_diff > 30
+                right_masks = right_diff > 30
+            
+            # Calculate areas for this batch
+            left_areas = cp.sum(left_masks, axis=(1, 2))
+            right_areas = cp.sum(right_masks, axis=(1, 2))
+            area_ratios = right_areas.astype(cp.float32) / (left_areas.astype(cp.float32) + 1e-6)
+            
+            # Find frames needing correction in this batch
+            needs_correction = (area_ratios < 0.5) | (area_ratios > 2.0)
+            
+            # Transfer batch results back to CPU and accumulate
+            area_ratios_all.extend(cp.asnumpy(area_ratios))
+            needs_correction_all.extend(cp.asnumpy(needs_correction))
+            
+            # Free GPU memory for this batch
+            del left_stack, right_stack, left_masks, right_masks
+            del left_areas, right_areas, area_ratios, needs_correction
+            cp._default_memory_pool.free_all_blocks()
+        
+        # CRITICAL: Release ALL CuPy memory back to system after validation
+        try:
+            # Force release of all GPU memory pools
+            cp._default_memory_pool.free_all_blocks()
+            cp._default_pinned_memory_pool.free_all_blocks()
+            
+            # Clear CuPy cache completely
+            cp.get_default_memory_pool().free_all_blocks()
+            cp.get_default_pinned_memory_pool().free_all_blocks()
+            
+            print(f"   CuPy memory pools cleared")
+        except Exception as e:
+            print(f"   Warning: Could not clear CuPy memory pools: {e}")
+        
+        correction_count = sum(needs_correction_all)
+        print(f"   GPU validation complete: {correction_count}/{total_frames} frames need correction")
+        
+        # Apply corrections using CPU results
+        validated_frames = []
+        for i, (needs_fix, ratio) in enumerate(zip(needs_correction_all, area_ratios_all)):
+            if i % 100 == 0:
+                print(f"   Processing validation results: {i}/{total_frames}")
+                
+            if needs_fix:
+                # Apply correction
+                corrected_frame = self._apply_stereo_correction(
+                    left_results[i], right_results[i], float(ratio)
+                )
+                validated_frames.append(corrected_frame)
+            else:
+                validated_frames.append(right_results[i])
+        
+        print("✅ VALIDATION: GPU stereo consistency check complete")
+        return validated_frames
+    
+    def _validate_stereo_consistency_cpu(self, 
+                                       left_results: List[np.ndarray], 
+                                       right_results: List[np.ndarray]) -> List[np.ndarray]:
+        """CPU fallback for stereo validation"""
+        print("   Using CPU validation (slower)")
         validated_frames = []
         
         for i, (left_frame, right_frame) in enumerate(zip(left_results, right_results)):
+            if i % 50 == 0:  # Progress every 50 frames
+                print(f"   CPU validation progress: {i}/{len(left_results)}")
+            
             # Simple validation: check if mask areas are similar
             left_mask_area = self._get_mask_area(left_frame)
             right_mask_area = self._get_mask_area(right_frame)
@@ -448,6 +635,7 @@ class VR180Processor(VideoProcessor):
             else:
                 validated_frames.append(right_frame)
         
+        print("✅ VALIDATION: CPU stereo consistency check complete")
         return validated_frames
     
     def _create_empty_masks_from_count(self, num_frames: int, frame_shape: tuple) -> List[np.ndarray]:
@@ -465,7 +653,26 @@ class VR180Processor(VideoProcessor):
         return empty_frames
     
     def _get_mask_area(self, frame: np.ndarray) -> float:
-        """Get mask area from processed frame"""
+        """Get mask area from processed frame using GPU acceleration"""
+        try:
+            import cupy as cp
+            
+            # Transfer to GPU
+            frame_gpu = cp.asarray(frame)
+            
+            if frame.shape[2] == 4:  # Alpha channel
+                mask_gpu = frame_gpu[:, :, 3] > 0
+            else:  # Green screen - detect non-background pixels
+                bg_color_gpu = cp.array(self.config.output.background_color)
+                diff_gpu = cp.abs(frame_gpu.astype(cp.float32) - bg_color_gpu).sum(axis=2)
+                mask_gpu = diff_gpu > 30  # Threshold for non-background
+            
+            # Calculate area on GPU and return as Python int
+            area = int(cp.sum(mask_gpu))
+            return area
+            
+        except ImportError:
+            # Fallback to CPU NumPy if CuPy not available
             if frame.shape[2] == 4:  # Alpha channel
                 mask = frame[:, :, 3] > 0
             else:  # Green screen - detect non-background pixels
@@ -489,6 +696,64 @@ class VR180Processor(VideoProcessor):
         # TODO: Implement proper stereo correction algorithm
         return right_frame
     
+    def _complete_inter_chunk_cleanup(self, chunk_idx: int):
+        """
+        Complete inter-chunk cleanup: Destroy all models to prevent memory accumulation
+        
+        This addresses the core issue where SAM2 and YOLO models (~15-20GB) 
+        persist between chunks, causing OOM when processing subsequent chunks.
+        """
+        print(f"🧹 INTER-CHUNK CLEANUP: Destroying all models after chunk {chunk_idx}")
+        
+        # 1. Completely destroy SAM2 model (15-20GB)
+        if hasattr(self, 'sam2_model') and self.sam2_model is not None:
+            self.sam2_model.cleanup()  # Call existing cleanup
+            
+            # Force complete destruction of the model
+            try:
+                # Reset the model's loaded state so it will reload fresh
+                if hasattr(self.sam2_model, '_model_loaded'):
+                    self.sam2_model._model_loaded = False
+                
+                # Clear any cached state
+                if hasattr(self.sam2_model, 'predictor'):
+                    self.sam2_model.predictor = None
+                if hasattr(self.sam2_model, 'inference_state'):
+                    self.sam2_model.inference_state = None
+                
+                print(f"   ✅ SAM2 model destroyed and marked for fresh reload")
+                
+            except Exception as e:
+                print(f"   ⚠️  SAM2 destruction warning: {e}")
+        
+        # 2. Completely destroy YOLO detector (400MB+)
+        if hasattr(self, 'detector') and self.detector is not None:
+            try:
+                # Force YOLO model to be reloaded fresh
+                if hasattr(self.detector, 'model') and self.detector.model is not None:
+                    del self.detector.model
+                    self.detector.model = None
+                    print(f"   ✅ YOLO model destroyed and marked for fresh reload")
+                    
+            except Exception as e:
+                print(f"   ⚠️  YOLO destruction warning: {e}")
+        
+        # 3. Clear CUDA cache aggressively
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()  # Wait for all operations to complete
+            print(f"   ✅ CUDA cache cleared")
+        
+        # 4. Force garbage collection
+        import gc
+        collected = gc.collect()
+        print(f"   ✅ Garbage collection: {collected} objects freed")
+        
+        # 5. Memory verification
+        self._print_memory_step(f"After complete inter-chunk cleanup (chunk {chunk_idx})")
+        
+        print(f"🎯 RESULT: Models will reload fresh for next chunk (prevents 15-20GB accumulation)")
+    
     def process_chunk(self, 
                      frames: List[np.ndarray], 
                      chunk_idx: int = 0) -> List[np.ndarray]:
@@ -548,6 +813,9 @@ class VR180Processor(VideoProcessor):
                 combined = {'left': left_frame, 'right': right_frame}
             combined_frames.append(combined)
         
+        # CRITICAL: Complete inter-chunk cleanup for independent processing too
+        self._complete_inter_chunk_cleanup(chunk_idx)
+        
         return combined_frames
     
     def save_video(self, frames: List[np.ndarray], output_path: str):