Synthetic 3D plankton dataset towards biovolume estimation

Abstract

Plankton biovolume is an important measure for different factors, such as water quality indicator. For that, it is important to monitor plankton, for which, plankton imaging is one of the predominant means utilized. Nonetheless, 2D images cannot fully represent 3D information and there is a scarcity of existing 2D-3D paired plankton dataset. A paired dataset like that is needed to train models that can predict 3D information from 2D plankton images, which is a significant step towards biovolume estimation. For this thesis, a synthetic dataset was built with paired 2D images, rendered from their 3D models, with their associated ground truth depth maps. The synthetic dataset comprises 672 pairs of images, with 21 different species. This dataset was then utilized to train three depth estimation model architectures and evaluated both on synthetic data quantitatively and real plankton images qualitatively. The three architectures evaluated were a simple CNN, U-Net, and Depth Anything V2. The scope of this work was relative depth estimation, which was to act as a stepping stone towards biovolume estimation. In terms of numerical results, Depth Anything V2 outperformed the other two overall, but the performance varied depending on the morphology of the test species, with rounded shapes yielding better results than spiny or elongated ones. Owing to the lack of depth ground truth for real IFCB images, this quantitative lead could not be confirmed qualitatively. Overall, depth estimation from synthetic plankton data is a feasible foundation as a step towards plankton biovolume estimation, but is limited by the small range of morphologies available and the synthetic-to-real domain gap.