Blazewild/Hav-Cocap_avcaps

Hav-Cocap AVCaps Dataset

This dataset contains the complete CoCap and Hav-Cocap implementations for audio-visual captioning.

Contents

• CoCap/: Original CoCap framework implementation
• Hav-Cocap/: Enhanced Hav-Cocap implementation with improvements
• requirements.txt: Python dependencies for the project

Project Structure

CoCap

• cocap/: Core modules for audio-visual captioning
  • data/: Dataset loading and preprocessing
  • modeling/: Model architectures and training
  • modules/: Audio encoder, CLIP, BEATs, etc.
  • utils/: Utility functions
• configs/: Configuration files for different datasets (MSRVTT, MSVD, VATEX)
• tools/: Training and evaluation scripts
• model_zoo/: Pre-trained model checkpoints

Hav-Cocap

• dataset/AVCaps/: AVCaps dataset with videos and captions
  • Train/Val/Test splits
  • Video files (240p H.264 format)
  • Caption annotations (JSON format)
• learn/: Educational notebooks and examples
• model/: Model architecture implementations

Installation

pip install -r requirements.txt

Dataset Details

The AVCaps dataset includes:

• Audio-visual caption pairs
• Multiple splits (train/val/test)
• Preprocessed video files
• JSON caption annotations

Usage

See the training scripts in CoCap/tools/ for training examples:

• train_avcaps.py: Main training script for AVCaps dataset
• train_net.py: General training script

Citation

If you use this dataset or code, please cite the original CoCap paper and this dataset.

License

MIT License - See LICENSE file for details