PaperlayerView exact revision on arXiv

Source-linked AI summary

AB3DMOT: A Baseline for 3D Multi-Object Tracking and New Evaluation Metrics

Xinshuo Weng, Jianren Wang, David Held, Kris Kitani

arXiv:2008.08063v1cs.CVcs.MAcs.RO

TL;DR

Because KITTI's standard evaluation assesses 3D tracking on the image plane, the paper pairs a simple 3D tracking baseline with direct 3D evaluation and threshold-integrated metrics. On KITTI, it reports strong 3D MOT performance and a runtime of 207.4 FPS excluding the detector, the fastest among modern 3D MOT systems.

  • Problem

    KITTI evaluation for 3D MOT uses the image plane, and a direct 3D evaluation tool is unavailable; practical efficiency and system simplicity also remain concerns in recent MOT work.

    Source: §1 · p. 1, §3 · p. 2

  • Method

    The system uses LiDAR-based 3D detections with a constant-velocity 3D Kalman filter and Hungarian association, alongside a direct 3D evaluation tool and metrics integrated across recall.

    Source: §2 · p. 1, §3 · p. 2, §4 · p. 3

  • Results

    On KITTI, the system establishes strong 3D MOT performance and reaches 207.4 FPS excluding the detector, the fastest among modern 3D MOT systems.

    Source: §6 · p. 4, §5 · p. 4

  • Takeaways & Limitations

    The authors intend the released system to serve as a baseline and the evaluation tool as a standard for future 3D MOT benchmarks.

    Source: §6 · p. 4

Abstract

from arXiv · show

3D multi-object tracking (MOT) is essential to applications such as autonomous driving. Recent work focuses on developing accurate systems giving less attention to computational cost and system complexity. In contrast, this work proposes a simple real-time 3D MOT system with strong performance. Our system first obtains 3D detections from a LiDAR point cloud. Then, a straightforward combination of a 3D Kalman filter and the Hungarian algorithm is used for state estimation and data association. Additionally, 3D MOT datasets such as KITTI evaluate MOT methods in 2D space and standardized 3D MOT evaluation tools are missing for a fair comparison of 3D MOT methods. We propose a new 3D MOT evaluation tool along with three new metrics to comprehensively evaluate 3D MOT methods. We show that, our proposed method achieves strong 3D MOT performance on KITTI and runs at a rate of $207.4$ FPS on the KITTI dataset, achieving the fastest speed among modern 3D MOT systems. Our code is publicly available at http://www.xinshuoweng.com/projects/AB3DMOT.

1 Introduction

The introduction frames MOT as important to applications such as autonomous driving, while noting that accuracy-focused progress has left practical efficiency and simplicity concerns.

  • Recent MOT progress has emphasized accuracy, while computational efficiency and system simplicity remain important practical factors.The introduction also notes that modern state-of-the-art methods typically require substantial computation, making real-time performance a challenge.Source: §1 · p. 1
  • KITTI car-class 2D MOTA rose from 57.03 to 84.04 in two years.Source: §1 · p. 1
  • Complex modern MOT systems can make it unclear which system component contributes most to performance.Source: §1 · p. 1

2 AB3DMOT: A Baseline for 3D MOT

AB3DMOT combines a 3D Kalman filter and Hungarian data association as a simple baseline, extending object state representation to 3D. The authors report strong KITTI results and fast runtime.

  • An off-the-shelf detector supplies LiDAR-based 3D detections, then a constant-velocity 3D Kalman filter and Hungarian algorithm handle state estimation and data association.Source: §2 · p. 1
  • Its filter state includes 3D location, size, velocity, and orientation, unlike image-plane state spaces in other filter-based MOT systems.Source: §2 · p. 2
  • The straightforward system achieves strong 3D MOT performance on KITTI and 207.4 FPS, the fastest speed among modern 3D MOT systems.Source: §2 · p. 2
  • The system also achieves competitive KITTI 2D MOT leaderboard performance after projecting its 3D tracks for evaluation, despite using no 2D input; the authors hypothesize that 3D tracking better resolves depth ambiguities.Source: §2 · p. 2, Figure 1 · p. 2

3 New 3D MOT Evaluation Tool

The section identifies image-plane evaluation as a gap for 3D MOT and describes a direct 3D extension of KITTI's evaluation tool.

  • KITTI's standard convention evaluates 3D tracking results after projection to the image plane, and a direct 3D MOT evaluation tool is unavailable.Source: §3 · p. 2
  • The proposed extension replaces 2D IoU with 3D IoU and matches 3D tracking results directly to 3D ground-truth trajectories.Source: §3 · p. 2
  • A successful match requires at least 0.25 3D IoU in the experiments; the authors hope the tool can standardize future 3D MOT evaluation.Source: §3 · p. 2

4 New MOT Evaluation Metrics

The section argues that single-threshold MOT metrics miss confidence-score robustness and proposes integral metrics to summarize performance across thresholds.

  • MOTA and MOTP omit tracked-object confidence, forcing users to choose a threshold that can vary by detector or dataset and hide performance at other thresholds.Source: §4 · p. 3
  • The proposed sAMOTA, AMOTA, and AMOTP summarize MOTA and MOTP performance across different thresholds.Source: §4 · p. 3
  • AMOTA and AMOTP integrate MOTA and MOTP over recall, approximated with 40 recall values; sAMOTA scales AMOTA to a 0%-100% range.Source: §4 · p. 3

5 Experiments

On the KITTI 3D MOT benchmark, the system outperforms modern 3D MOT systems across all reported metrics and achieves zero identity switches; without a GPU, it runs at 207.4 FPS excluding the 3D detector.

5 Experiments

  • Evaluation uses the KITTI val car subset because test-set 3D ground truth is unavailable, comparing against open-sourced systems including FANTrack and mmMOT.The proposed method and baselines requiring 3D detections use PointRCNN detections; 2D-input baselines use their 2D projections.Source: §5 · pp. 3–4
  • The system consistently outperforms modern 3D MOT systems across all reported metrics, with zero identity switches.The comparison uses the proposed 3D MOT evaluation tool and new metrics.Source: §5 · p. 4
  • 207.4 FPS on the KITTI val set, excluding the 3D detector and without a GPU, is the fastest speed among modern 3D MOT systems.Source: §5 · p. 4

6 Conclusion

The paper presents an accurate, simple, real-time baseline for online 3D MOT and a 3D evaluation tool with new metrics. KITTI experiments show strong tracking performance and fastest speed, and the authors hope the system and tool will support future work and benchmarks.

6 Conclusion

  • The proposed system is an accurate, simple, real-time baseline for online 3D MOT.Source: §6 · p. 4
  • The paper proposes a 3D MOT evaluation tool and new metrics for standardized future evaluation.Source: §6 · p. 4
  • KITTI experiments establish strong 3D MOT performance while achieving the fastest speed.The authors hope the system will serve as a baseline for further 3D MOT work and the evaluation tool as a future benchmark standard.Source: §6 · p. 4
Loading 2008.08063v1…