Deep Affinity Net: Instance Segmentation via Affinity

Abstract: Most of the modern instance segmentation approaches fall into two categories: region-based approaches in which object bounding boxes are detected first and later used in cropping and segmenting instances; and keypoint-based approaches in which individual instances are represented by a set of keypoints followed by a dense pixel clustering around those keypoints. Despite the maturity of these two paradigms, we would like to report an alternative affinity-based paradigm where instances are segmented based on densely predicted affinities and graph partitioning algorithms. Such affinity-based approaches indicate that high-level graph features other than regions or keypoints can be directly applied in the instance segmentation task. In this work, we propose Deep Affinity Net, an effective affinity-based approach accompanied with a new graph partitioning algorithm Cascade-GAEC. Without bells and whistles, our end-to-end model results in 32.4% AP on Cityscapes val and 27.5% AP on test. It achieves the best single-shot result as well as the fastest running time among all affinity-based models. It also outperforms the region-based method Mask R-CNN.