Data Augmentation for Fish Instance Segmentation Using Automatic Object Generation

Abstract

In the field of fisheries, the management of fishery resources is one of the important tasks. Instance segmentation can be applied to fish species and fish body length estimation can be implemented using instance segmentation technology, which is expected to improve the fish-length estimation accuracy compared to the general object detection task that estimates rectangular regions. However, the annotation cost of training data is an issue to achieve instance segmentation. In this study, we propose a method of data augmentation by automatically generating and placing objects using the 3D-CG software Blender. By combining real images and computer graphics, our method achieves data augmentation such as random posing of fish and automatic generation of annotation data for instance segmentation. In the experiment, we evaluated the effectiveness of Mask-RCNN for instance segmentation on the dataset with our data augmentation, and found that the model trained on the data automatically generated by the proposed method is effective for instance segmentation for real images.