Releases: lRomul/ball-action-spotting
Releases · lRomul/ball-action-spotting
Trained models
Initial release
Solution
The solution includes an efficient model architecture and a multi-stage training pipeline for spotting the instances of passes and drives occurring in soccer videos.
Key points:
- Efficient model architecture for extracting information from video data
- Multi-stage training pipeline (transfer learning, fine-tuning with long sequences)
- Fast video loading pipeline for training (GPU decoding, no need for preprocessing with extracting images)