TSM: Temporal Shift Module for Efficient Video Understanding

Ji Lin1    Chuang Gan2    Song Han1

Abstract

The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN's complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition. The code is available at: https://github. com/mit-han-lab/temporal-shift-module.

Paper

• TSM: Temporal Shift Module for Efficient Video Understanding
  Ji Lin, Chuang Gan, and Song Han
  in ICCV 2019
  arXiv
• Training Kinetics in 15 Minutes: Large-scale Distributed Training on Videos
  Ji Lin, Chuang Gan, and Song Han
  in NeurIPS 2019 Systems for ML workshop
  arXiv
• Media Converage: MIT News / MIT Technology Review / WIRED / Engadget
• Code on GitHub

BibTeX

@inproceedings{lin2019tsm,
  title={TSM: Temporal Shift Module for Efficient Video Understanding},
  author={Lin, Ji and Gan, Chuang and Han, Song},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2019}
}  

Live Demo

[NEW] We have updated an online demo of hand gesture recognition, running on NVIDIA Jetson Nano (99$) at 8 watts. The model uses MobileNetV2 as the backbone and inserted online TSM. It is compiled with TVM. It can run at more than 70 FPS on Nano (for the demo, the speed is delayed by the camera).

We build a live demo on online hand gesture recognition with our TSM model. It is useful for driving scenarios where we can communicate with computers using hand gestures.

The model uses MobileNetV2 as the backbone, and TSM is only inserted for the last 3 residual blocks so that we can reuse the features to facilitate online recognition.

The demo is running on an NVIDIA Jetson TX2 board. It can recognize around 15 videos (8 frames) per second.

Introduction Video

This is a video introducing our TSM framework.