Important

This repository was archived by NVIDIA Deep Learning Institute on 2026-06-23. It is no longer maintained, and issues or pull requests are not being reviewed. For current NVIDIA DLI open-source training materials, visit https://github.com/NVDLI. This repository is a historical fork. For ongoing development, refer to the upstream project.

TAO: A Large-Scale Benchmark for Tracking Any Object

[Paper] [Website]

Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, Deva Ramanan

Latest updates

• [2020.10.25]: Added docs for running single-object trackers. See here for how to add your own SOT tracker!
• [2020.07.10]: The ECCV challenge is now live at the MOTChallenge website! See here for more details.
• [2020.07.02]: TAO was accepted to ECCV '20 as a spotlight presentation!
• [2020.02.20]: We will be hosting a workshop and challenge at ECCV'20. See here for details.

Setup

1. Clone this repo

   git clone https://github.com/TAO-Dataset/tao
   

2. Install TAO toolkit:

   pip install git+https://github.com/TAO-Dataset/tao
   

Download dataset

See download instructions.

Challenge

We will be hosting a challenge at our ECCV '20 workshop. See here for details.

Evaluation

See evaluation information. Contains information on submitting to the challenge server.

Run baseline trackers

See tracker instructions.

Questions?

Please see the faqs to check if we've anticipated your question. If not, for questions about TAO usage or the challenge, please use this Google Group: https://groups.google.com/forum/#!forum/tao-dataset/

For bug reports regarding the toolkit, annotations, or image download, please file an issue in this repository.