Learning Equivariant Segmentation with Instance-Unique Querying Supplementary Material Wenguan Wang ReLER, AAII, University of Technology Sydney James Liang

SparseInst [2] is a fast segmenter that learns a sparse set of instance-aware queries and predicts instances in a one-to-one style without non-maximum suppression. As seen, with the help of our algorithm, the performance is significantly boosted to 36.7 and 37.7 AP with ResNet-50 [4] and ResNet-50-DCN [5] backbones, which are 2.0 and 2.3 higher than the baseline, respectively. On the top of SOLQ, we also test our algorithm on the strong backbone -- Swin [6] backbone. Additional Panoptic Segmentation T ask. To accommodate the algorithmic setting, the only modification that needs to be made is for the stuff classes.