Region Proposal Rectification Towards Robust Instance Segmentation of Biological Images

Top-down instance segmentation framework has shown its superiority in object detection compared to the bottom-up framework. While it is efficient addressing over-segmentation, top-down suffers from over-crop problem. However, a complete mask crucial for biological image analysis as delivers important morphological properties such shapes and volumes. In this paper, we propose region proposal rectification (RPR) module address challenging incomplete particular, offer progressive ROIAlign introduce neighbor information into series of ROIs gradually. The ROI features are fed an attentive feed-forward network (FFN) box regression. With additional information, proposed RPR shows significant improvement correction locations thereby exhibits favorable performances on three datasets state-of-the-art baseline methods. Experimental results demonstrate that effective both anchor-based anchor-free approaches, suggesting method can be applied general images. Code available ( https://github.com/qzhangli/RPR ).