A Novel Encoder and Label Assignment for Instance Segmentation

SparseInst, a recent lightweight instance segmentation network, achieves better balance between efficiency and precision. However, the information contained in the single-layer features output by the encoder is not rich enough and the label assignment strategy leads to imbalance between positive and negative samples. In order to further improve the instance segmentation performance, we propose LAIS network including a novel feature encoding module and a Multi-Step Hungarian matching strategy (MSH). By combining multi-scale feature extraction and inter-layer information fusion, the encoder outputs features with more detailed and comprehensive information. By performing multiple rounds of one-to-one Hungarian matching, MSH eliminates the imbalance and duplication during the sample allocation. Experiments show that LAIS is more accurate than SparseInst without significantly increasing parameters and GFLOPs. In particular, LAIS reached 33.8% AP on the COCO val, 1.0% higher than SparseInst when using the same ResNet-50 Backbone.