Tracking Reflected Objects: A Benchmark

Xiaoyu Guo, Pengzhi Zhong, Lizhi Lin, Hao Zhang, Ling Huang, Shuiwang Li; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 1756-1776

Visual tracking has made great progress recently, thanks to large-scale training datasets. These datasets have enabled the development of highly accurate tracking algorithms. However, most research focuses on tracking general objects, with less attention on specialized challenges like tracking reflected objects. Reflections can distort objects' appearances, making them harder to track, which is a critical issue in areas like autonomous driving, security, smart homes, and industrial settings where tracking reflections in mirrors or glass is important. To fill this gap, we introduce TRO, a benchmark designed for Tracking Reflected Objects. TRO contains 200 sequences with about 70,000 frames, all annotated with bounding boxes, aiming to foster the development of methods tailored for this challenging task. We tested 20 state-of-the-art trackers and found they struggle with the complexities of reflections. To set a better baseline, we propose HIP-HaTrack, a new tracker that uses hierarchical features and outperforms existing algorithms. We hope our benchmark and HIP-HaTrack will inspire further research on tracking reflected objects. The TRO and code are available at https://github.com/Guoxiaoyuy/HIP-HaTrack.