Underwater Object Tracking Benchmark and Dataset

Abstract

While there has been tremendous advancement in object tracking for open air visual data, much less work has been done for underwater object tracking. This is due to the low quality of underwater visual data. Underwater visual data suffers distortions in contrast and sharpness, as a result of refraction and absorption of light, and particles, which all vary dependent on the depth, color and nature of water. Although there currently exists several object tracking algorithms with proven record of high speed, precision and success rate, these algorithms work best for open air tracking, and considerably degrade in performance when tracking targets in underwater environments, as it is presented in this paper. The advancement made in open air tracking has been facilitated by availability of multiple benchmark and dataset. However, no such benchmark and dataset exist for underwater tracking, and this lack of data has hindered development of dedicated underwater tracking algorithms. In this paper, we present: a) the first underwater tracking benchmark dataset consisting of 32 videos, and a total of 24241 annotated frames, averaging 29.15 seconds and 757.53 frames per video, to help improve underwater tracking; and b) a comparative performance analysis of existing tracking algorithms in underwater environment as opposed to open air.