VIRTUALLY THROWING BENCHMARKS INTO THE OCEAN FOR DEEP SEA PHOTOGRAMMETRY AND IMAGE PROCESSING EVALUATION

Abstract. Vision in the deep sea is acquiring increasing interest from many fields as the deep seafloor represents the largest surface portion on Earth. Unlike common shallow underwater imaging, deep sea imaging requires artificial lighting to illuminate the scene in perpetual darkness. Deep sea images suffer from degradation caused by scattering, attenuation and effects of artificial light sources and have a very different appearance to images in shallow water or on land. This impairs transferring current vision methods to deep sea applications. Development of adequate algorithms requires some data with ground truth in order to evaluate the methods. However, it is practically impossible to capture a deep sea scene also without water or artificial lighting effects. This situation impairs progress in deep sea vision research, where already synthesized images with ground truth could be a good solution. Most current methods either render a virtual 3D model, or use atmospheric image formation models to convert real world scenes to appear as in shallow water appearance illuminated by sunlight. Currently, there is a lack of image datasets dedicated to deep sea vision evaluation. This paper introduces a pipeline to synthesize deep sea images using existing real world RGB-D benchmarks, and exemplarily generates the deep sea twin datasets for the well known Middlebury stereo benchmarks. They can be used both for testing underwater stereo matching methods and for training and evaluating underwater image processing algorithms. This work aims towards establishing an image benchmark, which is intended particularly for deep sea vision developments.