Image panoptic segmentation network based on Efficientnet

The complete perception and segmentation of images has become a research hotspot in the field of computer vision with the continuous development of the computer vision industry. This image segmentation technique needs the understanding of a scene which can neither figure out individual instances or different actors in the environment as well as the general background semantics. This newly emerging panopic segmentation problem in recent years is solved in this paper with a lightweight and efficient panoptic segmentation network, LEPSNet, which is based on Efficientnet. We designed this architecture based on a Two way PAN, with the ability of efficiently encoding and fusing semantically rich multi-scale features as a backbone network. A new semantic segmentation head is designed to capture fine context information in the background. Proposed is an improved separable convolutional module, and applied to the whole network. This module also results in adding it to Mask R-CNN, making the instance segmentation head. Experiments have demonstrated that the architecture can achieve a level of performance leadership while keeping the model lightweight. In particular, we achieved 65.0 PQ on Cityscapes and 50.5 PQ on Indian Driving Dataset (IDD).