2s-AGCN presents a novel approach to action recognition using skeletal data. The project leverages Graph Convolutional Networks (GCNs) with a two-stream architecture to capture both spatial and temporal information from human pose skeletons. By adapting the graph convolutions, the model effectively learns complex relationships between body joints, leading to improved action classification accuracy. The core problem addressed is accurate and robust action recognition from raw skeletal data.
This project introduces an adaptive GCN framework designed specifically for skeleton-based action recognition, addressing limitations of previous methods. It incorporates a two-stream approach to process spatial and temporal cues effectively. The code has been updated to support PyTorch 0.4 and later, with an additional model variant (AAGCN) for enhanced performance. The repository includes comprehensive data preprocessing and training scripts.
- Data Preprocessing: Provides scripts for downloading and preprocessing data from NTU-RGB+D and Skeleton-Kinetics datasets.
- Model Training: Includes PyTorch scripts for training the 2s-AGCN model with configurable parameters.
- Ensemble Testing: Offers a script to ensemble the outputs of joint and bone-based models for improved accuracy.
The project originated from a CVPR 2019 publication and has been updated to support recent PyTorch versions. The code includes clear instructions for data preparation, training, and testing. While it has a reasonable number of stars and forks, ongoing maintenance activity is not explicitly evident from the provided information. The documentation is sufficient for basic usage.
This repository is valuable for researchers and practitioners working on action recognition using skeletal data. It offers a well-documented and functional implementation of the 2s-AGCN model, enabling experimentation and replication of the research results. It provides a good starting point for incorporating graph convolutional networks into action recognition systems, particularly those dealing with skeleton inputs.