The paper proposes a novel concept of path consistency for self-supervised multi-object tracking. The key idea is that to track an object through frames, the model can obtain multiple different association results by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent.
Based on this rationale, the authors generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss (PCL) that enforces the association results to be consistent across different observation paths. This allows the model to learn robust object matching over both short and long temporal distances without using any ground-truth object identity supervision.
The authors demonstrate that their method outperforms existing unsupervised methods with consistent margins on various evaluation metrics across three tracking datasets (MOT17, PersonPath22, KITTI), and even achieves performance close to supervised methods. Extensive ablation studies validate the effectiveness of the proposed path consistency loss in learning long-distance object matching and improving tracking performance under challenging scenarios like occlusion.
إلى لغة أخرى
من محتوى المصدر
arxiv.org
الرؤى الأساسية المستخلصة من
by Zijia Lu,Bin... في arxiv.org 04-09-2024
https://arxiv.org/pdf/2404.05136.pdfاستفسارات أعمق