Joint COCO and Mapillary Workshop at ICCV 2019: Panoptic Segmentation Challenge Track Technical Report: Explore Context Relation for Panoptic Segmentation

This technical report presents an effective method for the panoptic segmentation. For the stuff segmentation, we propose a parallel attention module and multi-stage context branch, which take advantage of the context relations. Besides, we propose a Top-K normalization method for individual model and an adaptive scale method for the model ensemble. For the instance segmentation, we adopt a classic two-pass pipeline. Our ensemble model achieves 56.43 mIOU on the stuff segmentation validation set. By fusing the stuff segmentation and the instance segmentation result, we obtain 54.5 in PQ on the test-dev set.