TransSnake: Contour-Based Wild Animal Instance Segmentation Using Contour Transformer
Tong Liu, Zheng Zhang, Zedong Wu, Bochuan Zheng · 2024
Wild animal instance segmentation technology plays a significant role in wild animal research, enabling automatic wild animal information extraction, improving data processing speed, and reducing researcher burden. Inspired by the Snake algorithm and Transformer, we propose TransSnake, a new contour-based instance segmentation model. Its process included three stages: initial contour, coarse contour, and refined contour. In the coarse contour stage, we introduce a multi-scale fusion module to fuse multi-scale contour information, expand the receptive field, and generate a higher-quality coarse contour. The obtained coarse contour lays a solid foundation for subsequent contour refinement. However, relying solely on local information for contour deformation can result in substantial prediction errors. Therefore, we adopt the contour transformer in the refined contour stage to capture global information among contour points. This module uses global information to guide contour deformation and improve segmentation accuracy. Experimental results illustrate the outstanding performance of our method on the LoTE-Animal dataset and validate the feasibility of contour-based instance segmentation methods in wild animal segmentation. Meanwhile, the excellent performance is also shown from our approach on the SBD dataset.