Improving Panoptic Segmentation in Occlusion with a Local-aware Decoder

Panoptic segmentation, a unifying computer vision task combining semantic and instance segmentation, plays a crucial role in autonomous driving and robot navigation applications. With advancements in deep learning technology, particularly the introduction of the Transformer model, panoptic segmentation achieved significant progress. However, existing methods still struggle with performance in occluded and detailed regions. In this network, a local aware decoder has been designed to replace the Transformer decoder, guiding the network to concentrate on local areas. The proposed network improves segmentation accuracy, reduces the model’s size, and lowers computational costs. The experimental results indicate that our proposed method is effective, attaining a competitive PQ score of 63.0 on the Cityscapes dataset.