Learning panoptic segmentation through feature discriminability
作者：
Highlights：
• Different requirements of sub-tasks lead to conflicts of feature discriminability.
• Separating shared features can enhance feature discriminability of each sub-task.
• Region refinement module resolves conflict of feature discriminative between regions.
• We report state-of-the-art panoptic segmentation performance on two challenge datasets.
摘要
•Different requirements of sub-tasks lead to conflicts of feature discriminability.•Separating shared features can enhance feature discriminability of each sub-task.•Region refinement module resolves conflict of feature discriminative between regions.•We report state-of-the-art panoptic segmentation performance on two challenge datasets.
论文关键词：Panoptic segmentation,Feature discriminability,Region refinement
论文评审过程：Received 26 October 2020, Revised 18 June 2021, Accepted 8 August 2021, Available online 18 August 2021, Version of Record 23 August 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108240
