A Systematic Approach to Synthesize Underwater Images Benchmark Dataset and Beyond
Xiaodong Liu, Ben M. Chen · 2019
Underwater imaging suffers from a severe light attenuation. Images captured underwater often appear color cast with limited visibility, which may hinder the performance for underwater vision tasks. To handle this, many underwater image enhancement methods are proposed but their results are evaluated on different datasets. The lack of a large diverse dataset to efficiently evaluate these enhancement methods motivates this work. In this paper, we propose a systematic approach to synthesize a large diversity of underwater images as benchmark dataset. This dataset is generated on the basis of NYU-V2 indoor RGB-D dataset. The intensity of underwater ambient light is simulated based on statistic law and the attenuation coefficients are carefully selected. 1449 indoor images function as the ground-truth and 10 types of underwater scenes are synthesized with large depth range. Based on the benchmark, many image enhancement methods are sufficiently evaluated. Besides, the dataset are also beneficial to develop data-driven-based methods. The synthesized dataset will be made publicly available.