3D-ResNets-PyTorch implements 3D Residual Networks (3D ResNets) for spatiotemporal action recognition. The project focuses on achieving improved performance in action recognition tasks by leveraging 3D convolutional neural networks. It addresses the challenge of capturing both spatial and temporal information from video data, a key aspect for accurate action understanding. This implementation supports training, fine-tuning, and testing on various datasets commonly used for action recognition research.
This project offers pre-trained 3D ResNet models on multiple datasets, enabling faster experimentation and serving as a strong baseline. The code supports distributed training for increased scalability. The repository includes comprehensive instructions for data preparation and usage, making it accessible to researchers. It also offers flexibility in training configuration with options to change model depth, pretraining classes, and GPU utilization.
- Data Loading: Supports datasets including Kinetics, Moments in Time, ActivityNet, UCF-101, and HMDB-51.
- Pre-trained Models: Provides pre-trained 3D ResNet models on Kinetics-700, Moments in Time, STAIR-Actions, and merged datasets.
- Training & Fine-tuning: Includes scripts for training and fine-tuning 3D ResNet models.
- Data Preparation: Offers scripts to convert video data (Avi) to JPG format and generate annotation files in JSON format.
- Distributed Training: Supports training on multiple GPUs for faster training times.
- Configuration Flexibility: Allows customization of various training parameters such as model depth, pretraining classes, and dataset paths.
- Clear Instructions: Provides detailed instructions for data preparation, model training, and usage of pre-trained models.
The project is a finalized implementation with a well-defined structure and comprehensive documentation. The code is actively maintained and includes updates for supporting newer PyTorch versions and distributed training. The availability of pre-trained models and detailed instructions suggests a stable and reliable codebase.
This repository is valuable for researchers and practitioners involved in action recognition. It provides readily usable code, pre-trained models, and comprehensive instructions for training and evaluation. Users can leverage this project to quickly prototype and evaluate 3D ResNet models on their own datasets, benefiting from its well-structured implementation and established practices.