Tsnake: A Time-Embedded Recurrent Contour-Based Instance Segmentation Model

Instance segmentation is a critical task in computer vision. It is widely applied in autonomous driving and medical image analysis. Contour-based methods have been studied for their elegant structures, compact representation, and low inference time. In existing work, contour-based methods apply iterative reformation to fit the object boundary, however, the modules are independent in each loop, which enlarges the model sizes. In this paper, we introduce the TSnake. It is a novel time-embedded driven recurrent contour-based instance segmentation model. At each iteration, the refinement module adjusts the contour with the help of semantic-rich features and the time step. Moreover, a dilated module is applied to well adopt the local contextual information. We also modify the loss function to well balance the trade-off between over-smoothing and over-sensitivity of the contour. The proposed TSnake method surpasses the contour-based state-of-the-art instance segmentation and can be performed in real time.