2PN: A Unified Panoptic Segmentation Network with Attention Module

Abstract

Comprehensive and accurate surveillance of the environment forms the basis of secure Internet of things (IoTs), the threats can be observed, and the AI services of IoT systems can be preserved. Panoptic segmentation is an efficient and popular approach for environmental surveillance based on images captured by smart sensing devices. This approach can jointly detect stuffs and things within an image and feed subsequent tasks like image detection. So far, there are many methods for panoptic segmentation which focus on extracting sophisticated visual features for segmentation. However, these efforts are both heavy on their workload and cannot clearly distinguish essential features useful for surveillance in an open environment. Therefore, this paper proposes a novel deep learning model 2PN for panoptic segmentation. The model includes a 2-way pyramid network and an attention module to learn in a more concentrated and reasonable way which enhances the feature extraction part. It strikes a balance between the computing complexity and the power of model capability. Finally, 2PN (2-way pyramid network) results are reflected on the Cityscapes dataset.