Skip to content

Optional Steps

Fine-tune YOLO for Your Dataset

If you wish to use YOLO to automatically generate detections, you may want to fine-tune your YOLO model for your dataset using the YOLO training notebook.

cvat2ultralytics

Convert CVAT annotations to Ultralytics YOLO dataset format for training custom models.

Usage: 

cvat2ultralytics --video path_to_videos --annotation path_to_annotations --dataset dataset_name [--skip skip_frames]

Source: src/kabr_tools/cvat2ultralytics.py

Additional Utility Tools

player

Interactive player for tracking and behavior observation with visualization capabilities.

Usage: 

player --folder path_to_folder [--save] [--imshow]

Player Output

Figure: Example player.py output showing tracking visualization.

Source: src/kabr_tools/player.py

cvat2slowfast

Convert CVAT annotations to the dataset in Charades format for use with SlowFast video understanding models.

Usage: 

cvat2slowfast --miniscene path_to_mini_scenes --dataset dataset_name --classes path_to_classes_json [--old2new path_to_old2new_json] [--no_images]

Source: src/kabr_tools/cvat2slowfast.py

Helper Scripts

Several utility scripts are available in the helper_scripts directory:

Annotate Mini-scenes

Scripts for batch processing and annotation of mini-scenes:

  • launch_job.py - Job launcher for batch processing.
  • run.sh - Shell script for running annotation tasks.

See README for detailed usage instructions.

Video Processing Utilities

  • downgrade.sh - Reduce video file sizes for CVAT compatibility.
  • rename.sh - Batch rename files with consistent naming conventions.

Advanced Configuration

Custom Model Training

For training custom behavior recognition models, you'll need:

  1. Annotated mini-scenes from your specific species/environment.
  2. Behavior ethogram defining your classification categories.
  3. Training configuration adapted to your dataset size and complexity.

GPU Requirements

Most tools support both CPU and GPU processing:

  • CPU mode: Slower but works on any system.
  • GPU mode: Significantly faster, recommended for large datasets.
  • Multi-GPU: Supported for distributed processing of large video collections.

Batch Processing

For processing large video collections:

  1. Use the provided shell scripts for batch operations.
  2. Consider using job scheduling systems (SLURM, PBS) for cluster environments.
  3. Monitor disk space requirements for intermediate files.

Customization Options

The modular design allows for customization at multiple levels:

  • Detection models: Swap YOLO versions or use custom trained models.
  • Tracking algorithms: Modify tracking parameters for different scenarios.
  • Behavior models: Train species-specific or behavior-specific classifiers.
  • Analysis pipelines: Adapt visualization and metrics to research questions.

Performance Optimization

Tips for optimizing performance:

  • Use appropriate video resolution for your analysis needs.
  • Batch process multiple videos simultaneously when possible.
  • Utilize GPU acceleration for model inference.
  • Consider video compression for storage efficiency.