UTB180: A High-quality Benchmark for Underwater Tracking

Basit Alawode, Yuhang Guo, Mehnaz Ummar, Naoufel Werghi, Jorge Dias, Ajmal Mian, Sajid Javed; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 3326-3342

Deep learning methods have demonstrated encouraging performance on open-air visual object tracking (VOT) benchmarks, however, their strength remains unexplored on underwater video sequences due to the lack of challenging underwater VOT benchmarks. Apart from the challenges of open-air tracking, videos captured in underwater environments pose additional challenges for tracking such as low visibility, poor video quality, distortions in sharpness and contrast, reflections from suspended particles, and non-uniform lighting. We propose a new underwater tracking benchmark dataset (UTB180) consisting of 180 sequences to facilitate the development of underwater deep trackers. The sequences in UTB180 are selected from both underwater natural and online sources with over 58,000 annotated frames. Video-level attributes are also provided to facilitate the development of robust trackers for specific challenges. We benchmark 15 existing pre-trained State-Of-The-Art (SOTA) trackers on UTB180 and compare their performance on another publicly available underwater benchmark. The trackers consistently perform worse on UTB180 showing that it poses more challenging scenarios. Moreover, we show that fine-tuning five high-quality SOTA trackers on UTB180 still does not sufficiently boost their tracking performance. Our experiments show that the UTB180 sequences pose a major burden on the SOTA trackers as compared to their open-air tracking performance. The performance gap reveals the need for a dedicated end-to-end underwater deep tracker that takes into account the inherent properties of underwater environments. We believe that our proposed dataset will be of great value to the tracking community in advancing the state of the art in underwater VOT. Our dataset is publicly available on Kaggle.