zyg11 / alphavideo

Vision toolbox for video related tasks including action recognition, multi-object tracking.

Introduction

AlphaVideo is an open-sourced video understanding toolbox based on PyTorch covering multi-object tracking and action detection. In AlphaVideo, we released the first one-stage multi-object tracking (MOT) system TubeTK that can achieve 66.9 MOTA on MOT-16 dataset and 63 MOTA on MOT-17 dataset. For action detection, we released an efficient model AlphAction, which is the first open-source project that achieves 30+ mAP (32.4 mAP) with single model on AVA dataset.

Quick Start

pip

Run this command:

pip install alphavideo

from source

Clone repository from github:

git clone https://github.com/Alpha-Video/AlphaVideo.git alphaVideo
cd alphaVideo

Setup and install AlphaVideo:

pip install .

Features & Capabilities

• Multi-Object Tracking

  For this task, we provide the TubeTK model which is the official implementation of paper "TubeTK: Adopting Tubes to Track Multi-Object in a One-Step Training Model (CVPR2020, oral)." Detailed training and testing script on MOT-Challenge datasets can be found here.

  
  • Accurate end-to-end multi-object tracking.
  • Do not need any ready-made image-level object deteaction models.
  • Pre-trained model for pedestrian tracking.
  • Input: Frame list; video.
  • Output: Videos decorated by colored bounding-box; Btube lists.
  • For details usages, see our docs.

• Action recognition

  For this task, we provide the AlphAction model as an implementation of paper "Asynchronous Interaction Aggregation for Action Detection".

  
  • Accurate and efficient action detection.
  • Pre-trained model for 80 atomic action categories defined in AVA.
  • Input: Video; camera.
  • Output: Videos decorated by human boxes, attached with corresponding action predictions.
  • For details usages, see our docs.

Paper and Citations

@inproceedings{pang2020tubeTK,
  title={TubeTK: Adopting Tubes to Track Multi-Object in a One-Step Training Model},
  author={Pang, Bo and Li, Yizhuo and Zhang, Yifan and Li, Muchen and Lu, Cewu}
  booktitle={CVPR},
  year={2020}
}

@article{tang2020asynchronous,
  title={Asynchronous Interaction Aggregation for Action Detection},
  author={Tang, Jiajun and Xia, Jin and Mu, Xinzhi and Pang, Bo and Lu, Cewu},
  journal={arXiv preprint arXiv:2004.07485},
  year={2020}
}

Maintainers

This project is open-sourced and maintained by Machine Vision and Intelligence Group (MVIG) in Shanghai Jiao Tong University.