VastTrack: Vast Category Visual Object Tracking

• DOI: 10.48550/arxiv.2403.03493
• Publication Date: 2024-03-06
• Authors: Liang Peng; Junyuan Gao; Xinran Liu; Weihong Li; Shaohua Dong; Zhipeng Zhang; Heng Fan; Libo Zhang
• Keywords: Tracking (education)

In this paper, we introduce a novel benchmark, dubbed VastTrack, towards facilitating the development of more general visual tracking via encompassing abundant classes and videos. VastTrack possesses several attractive properties: (1) Vast Object Category. particular, it covers target objects from 2,115 classes, largely surpassing object categories existing popular benchmarks (e.g., GOT-10k with 563 LaSOT 70 categories). With such vast expect to learn tracking. (2) Larger scale. Compared current benchmarks, offers 50,610 sequences 4.2 million frames, which makes date largest benchmark regarding number videos, thus could benefit training even powerful trackers in deep learning era. (3) Rich Annotation. Besides conventional bounding box annotations, also provides linguistic descriptions for The rich annotations enables both vision-only vision-language To ensure precise annotation, all videos are manually labeled multiple rounds careful inspection refinement. understand performance provide baselines future comparison, extensively assess 25 representative trackers. results, not surprisingly, show significant drops compared those on datasets due lack diverse scenarios training, efforts required improve Our evaluation results will be made publicly available https://github.com/HengLan/VastTrack.