A Bottom-Up Approach for Multi Components in Contour-Based Instance Segmentation

Contour-based instance segmentation methods have been studied for their elegant structure and flexibility. However, these methods are hard to deal with multiple component objects since the restriction of contours. Usually, top-down method is used with the help of two detectors, but the computation time increases for the additional detector and filtering bounding boxes. In this paper, we derive the CCNet, a novel bottom-up approach to combining components into objects. We apply the affinity of each contour with self-attention to avoid double detecting and NMS. Experiments on the Cityscapes dataset show the proposed CCNet achieves comparable accuracy to two-detector approaches while maintaining inference speed close to single-component models. Further-more, our approach can be added to each kind of contour-based segmentation model. The CCNet provides an effective solution for multi-component instance segmentation that balances accuracy and efficiency.