EnzoLiang / PaddleVideo

Comprehensive, latest, and deployable video deep learning algorithm, including video recognition, action localization, and temporal action detection tasks. It's a high-performance, light-weight codebase provides practical models for video understanding research and application

简体中文 | English

PaddleVideo

Introduction

PaddleVideo is a toolset for video recognition, action localization, and spatio temporal action detection tasks prepared for the industry and academia. This repository provides examples and best practice guildelines for exploring deep learning algorithm in the scene of video area. We devote to support experiments and utilities which can significantly reduce the "time to deploy". By the way, this is also a proficiency verification and implementation of the newest PaddlePaddle 2.0 in the video field.

If you think this repo is helpful to you, welcome to star us~ ⭐

Features

• Various dataset and models PaddleVideo supports more datasets and models, including Kinectics400, UCF101, YoutTube8M, NTU-RGB+D datasets, and video recognition models, such as TSN, TSM, SlowFast, TimeSformer, AttentionLSTM, ST-GCN and action localization model, like BMN.

• Higher performance PaddleVideo has built-in solutions to improve accuracy on recognition models. PP-TSM, which is based on the standard TSM, already archive the best performance in the 2D recognition network, has the same size of parameters but improve the Top1 Acc to 76.16%.

• Faster training strategy PaddleVideo suppors faster training strategy, such as AMP training, Distributed training, Multigrid method for Slowfast, OP fusion method, Faster reader and so on.

• Deployable PaddleVideo is powered by the Paddle Inference. There is no need to convert the model to ONNX format when deploying it, all you want can be found in this repository.

• Applications PaddleVideo provides some interesting and practical projects that are implemented using video recognition and detection techniques, such as FootballAction and VideoTag.

Overview of the performance

Field	Model	Dataset	Metrics	ACC%
action recognition	PP-TSM	Kinetics-400	Top-1	76.16
action recognition	PP-TSN	Kinetics-400	Top-1	75.06
action recognition	AGCN	FSD-10	Top-1	90.66
action recognition	ST-GCN	FSD-10	Top-1	86.66
action recognition	TimeSformer	Kinetics-400	Top-1	77.29
action recognition	SlowFast	Kinetics-400	Top-1	75.84
action recognition	TSM	Kinetics-400	Top-1	71.06
action recognition	TSN	Kinetics-400	Top-1	69.81
action recognition	AttentionLSTM	Youtube-8M	Hit@1	89.0
action detection	BMN	ActivityNet	AUC	67.23

Changelog

release/2.1 was released in 20/05/2021. Please refer to release notes for details.

Community

• Scan the QR code below with your Wechat and reply "video", you can access to official technical exchange group. Look forward to your participation.

Applications

• VideoTag: 3k Large-Scale video classification model

• FootballAction: Football action detection model

Tutorials and Docs

• Tutorials and Slides
  • 2021.01
  • Summarize of video understanding
• Quick Start
  • Install
  • Start
• Project design
  • Modular design
  • Configuration design
• Model zoo
  • recognition
    • TimeSformer
    • Attention-LSTM
    • TSN
    • TSM
    • PP-TSM
    • PP-TSN
    • SlowFast
  • Localization
    • BMN
  • Skeleton-based action recognition
    • ST-GCN
    • AGCN
  • Spatio temporal action detection
    • Coming Soon!
  • ActBERT: Learning Global-Local Video-Text Representations
    • Coming Soon!
• Practice
  • Higher performance PP-TSM
  • Accelerate training
  • Deployment
• Others
  • Benchmark
  • Tools

License

PaddleVideo is released under the Apache 2.0 license.

Contributing

This poject welcomes contributions and suggestions. Please see our contribution guidelines.

• Many thanks to mohui37 for contributing the code for prediction.