streaming part1

config-streaming-runpod.yaml

@@ -0,0 +1,70 @@
+# VR180 Streaming Configuration for RunPod
+# Optimized for A6000 (48GB VRAM) or similar cloud GPUs
+
+input:
+  video_path: "/workspace/input_video.mp4"  # Update with your input path
+  start_frame: 0  # Resume from checkpoint if auto_resume is enabled
+  max_frames: null  # null = process entire video, or set a number for testing
+
+streaming:
+  mode: true  # True streaming - no chunking!
+  buffer_frames: 10  # Small buffer for correction lookahead
+  write_interval: 1  # Write every frame immediately
+
+processing:
+  scale_factor: 0.5  # 0.5 = 4K processing for 8K input (good balance)
+  adaptive_scaling: true  # Dynamically adjust scale based on GPU load
+  target_gpu_usage: 0.7  # Target 70% GPU utilization
+  min_scale: 0.25  # Never go below 25% scale
+  max_scale: 1.0  # Can go up to full resolution if GPU allows
+
+detection:
+  confidence_threshold: 0.7  # Person detection confidence
+  model: "yolov8n"  # Fast model suitable for streaming (n/s/m/l/x)
+  device: "cuda"
+
+matting:
+  sam2_model_cfg: "sam2.1_hiera_l"  # Use large model for best quality
+  sam2_checkpoint: "segment-anything-2/checkpoints/sam2.1_hiera_large.pt"
+  memory_offload: true  # Critical for streaming - offload to CPU when needed
+  fp16: true  # Use half precision for memory efficiency
+  continuous_correction: true  # Periodically refine tracking
+  correction_interval: 300  # Correct every 5 seconds at 60fps
+
+stereo:
+  mode: "master_slave"  # Left eye detects, right eye follows
+  master_eye: "left"  # Which eye leads detection
+  disparity_correction: true  # Adjust for stereo parallax
+  consistency_threshold: 0.3  # Max allowed difference between eyes
+  baseline: 65.0  # Interpupillary distance in mm
+  focal_length: 1000.0  # Camera focal length in pixels
+
+output:
+  path: "/workspace/output_video.mp4"  # Update with your output path
+  format: "greenscreen"  # "greenscreen" or "alpha"
+  background_color: [0, 255, 0]  # RGB for green screen
+  video_codec: "h264_nvenc"  # GPU encoding (or "hevc_nvenc" for better compression)
+  quality_preset: "p4"  # NVENC preset (p1-p7, higher = better quality)
+  crf: 18  # Quality (0-51, lower = better, 18 = high quality)
+  maintain_sbs: true  # Keep side-by-side format with audio
+
+hardware:
+  device: "cuda"
+  max_vram_gb: 40.0  # Conservative limit for 48GB GPU
+  max_ram_gb: 48.0  # RunPod container RAM limit
+
+recovery:
+  enable_checkpoints: true  # Save progress for resume
+  checkpoint_interval: 1000  # Save every ~16 seconds at 60fps
+  auto_resume: true  # Automatically resume from last checkpoint
+  checkpoint_dir: "./checkpoints"
+
+performance:
+  profile_enabled: true  # Track performance metrics
+  log_interval: 100  # Log progress every 100 frames
+  memory_monitor: true  # Monitor RAM/VRAM usage
+
+# Usage:
+# 1. Update input.video_path and output.path
+# 2. Adjust scale_factor based on your GPU (0.25 for faster, 1.0 for quality)
+# 3. Run: python -m vr180_streaming config-streaming-runpod.yaml