UniInst: Unique representation for end-to-end instance segmentation

Existing instance segmentation methods have achieved impressive performance but still suffer from a common dilemma: redundant representations (e.g., multiple boxes, grids, and anchor points) are inferred for one instance, which leads to duplicated predictions. Thus, mainstream usually rely on hand-designed non-maximum suppression (NMS) post-processing step select the optimal prediction result, consequently hindering end-to-end training. To address this issue, we propose box-free NMS-free framework, dubbed UniInst, yields only unique representation each instance. Specifically, design an instance-aware one-to-one assignment scheme, named Only Yield One Representation (OYOR). It dynamically assigns according matching quality between predictions ground truths. Then, novel re-ranking strategy is elegantly integrated into framework misalignment classification score mask quality, enabling learned be more discriminative. With these techniques, our first FCN-based achieves competitive performance, e.g., 39.0 AP using ResNet-50-FPN 40.2 ResNet-101-FPN COCO test-dev. Moreover, proposed method robust occlusion scenes because of non-dependent box NMS. outperforms baselines by remarkable heavily-occluded OCHuman benchmark. Code available at https://github.com/b03505036/UniInst.