Lightweight dual-branch network for vehicle exhausts segmentation

Visual vehicle exhausts segmentation is a novel and highly challenging task. In this paper, we introduce a lightweight dual-branch vehicle exhausts segmentation network to quickly and accurately infer segmentation masks from multi-interference traffic scenes. Firstly, we propose an encoder-decoder architecture with lightweight residual modules, which is divided into a deep branch for global prediction and a shallow branch for spatial details. Secondly, pyramid attention structure and skip modules are used to expand the receiving range and integrate multi-scale features. Finally, we advance a fusion network to merge the results of two branches so that the entire model can be easily trained end-to-end. To replace the complicated manual annotation, we employ dynamic fluid simulation and computer graphics technology to generate synthetic vehicle exhausts datasets VED. Comprehensive experiments on our synthetic and real datasets demonstrate that the proposed network outperforms existing segmentation networks in terms of speed and accuracy trade-off. Vehicle exhausts segmentation results on real videos are also appealing.