sAP — Code for Towards Streaming Perception

ECCV Best Paper Honorable Mention Award

Feb 2021: Announcing the Streaming Perception Challenge (CVPR 2021)!

This repo contains code for our ECCV 2020 paper (Towards Streaming Perception). sAP stands for streaming Average Precision.

The dataset used in this project (Argoverse-HD) can be found on the project page.

Apr 2021: Note that the code has been updated to match the setting in the streaming perception challenge (mostly about working with newer dependencies). To reproduce the numbers in our paper, please check out the ECCV paper branch.

Contents

• Offline detection
• Streaming (real-time online) detection
• Streaming tracking & forecasting
• Simulated streaming detection, tracking, & forecasting
• Simulated streaming detection, tracking, & forecasting with infinite GPUs
• Meta-detector Streamer
• Streaming evaluation
• Single-frame schedule simulator
• Helper functions for visualization

Getting started

1. Follow the instructions here to download and set up the dataset.
2. Follow the instructions here to install the dependencies.
3. Check out the examples to run various tasks in exp/*. The documentation for these tasks can be found here.

Citation

If you use the code or the data for your research, please cite the paper:

@article{Li2020StreamingP,
  title={Towards Streaming Perception},
  author={Li, Mengtian and Wang, Yuxiong and Ramanan, Deva},
  journal={ECCV},
  year={2020}
}

Acknowledgement

We would like to thank the mmdetection team for implementing so many different detectors in a single awesome repo with a unified interface! This greatly reduced our efforts to evaluate different detectors under our streaming setting.