/qd-3dt

Official implementation of Monocular Quasi-Dense 3D Object Tracking

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Monocular Quasi-Dense 3D Object Tracking

Monocular Quasi-Dense 3D Object Tracking (QD-3DT) is an online framework detects and tracks objects in 3D using quasi-dense object proposals from 2D images.

Monocular Quasi-Dense 3D Object Tracking,
Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, Min Sun,
arXiv technical report (arXiv 2103.07351) Project Website (QD-3DT)

@article{Hu2021QD3DT,
    author = {Hu, Hou-Ning and Yang, Yung-Hsu and Fischer, Tobias and Yu, Fisher and Darrell, Trevor and Sun, Min},
    title = {Monocular Quasi-Dense 3D Object Tracking},
    journal = {ArXiv:2103.07351},
    year = {2021}
}

Abstract

A reliable and accurate 3D tracking framework is essential for predicting future locations of surrounding objects and planning the observer’s actions in numerous applications such as autonomous driving. We propose a framework that can effectively associate moving objects over time and estimate their full 3D bounding box information from a sequence of 2D images captured on a moving platform. The object association leverages quasi-dense similarity learning to identify objects in various poses and viewpoints with appearance cues only. After initial 2D association, we further utilize 3D bounding boxes depth-ordering heuristics for robust instance association and motion-based 3D trajectory prediction for re-identification of occluded vehicles. In the end, an LSTM-based object velocity learning module aggregates the long-term trajectory information for more accurate motion extrapolation. Experiments on our proposed simulation data and real-world benchmarks, including KITTI, nuScenes, and Waymo datasets, show that our tracking framework offers robust object association and tracking on urban-driving scenarios. On the Waymo Open benchmark, we establish the first camera-only baseline in the 3D tracking and 3D detection challenges. Our quasi-dense 3D tracking pipeline achieves impressive improvements on the nuScenes 3D tracking benchmark with near five times tracking accuracy of the best vision-only submission among all published methods.

Main results

3D tracking on nuScenes test set

We achieved the best vision-only submission

AMOTA AMOTP
21.7 1.55

3D tracking on Waymo Open test set

We established the first camera-only baseline on Waymo Open

MOTA/L2 MOTP/L2
0.0001 0.0658

2D vehicle tracking on KITTI test set

MOTA MOTP
86.44 85.82

Installation

Please refer to INSTALL.md for installation and to DATA.md dataset preparation.

Get Started

Please see GETTING_STARTED.md for the basic usage of QD-3DT.

MODEL ZOO

Please refer to MODEL_ZOO.md for reproducing the results on varients of benchmarks

Contact

This repo is currently maintained by Hou-Ning Hu (@eborboihuc), Yung-Hsu Yang (@RoyYang0714), and Tobias Fischer (@tobiasfshr).

License

This work is licensed under BSD 3-Clause License. See LICENSE for details. Third-party datasets and tools are subject to their respective licenses.

Acknowledgements

We thank Jiangmiao Pang for his help in providing the qdtrack codebase in mmdetection. This repo uses py-motmetrics for MOT evaluation, waymo-open-dataset for Waymo Open 3D detection and 3D tracking task, and nuscenes-devkit for nuScenes evaluation and preprocessing.