TBANet: Three Branches with Asymmetric Convolutions and Pooling for Real-time Semantic Segmentation

Computational pressure limits the application of semantic segmentation in road scenes. We propose the lightweight network TBANet that makes a trade-off between accuracy and speed to deal with this problem. TBANet takes advantage of three branches (a fine-grained branch, an adaptive branch, and a coarse-grained branch) to obtain detailed and abstract information and adaptively supplement useful features for semantic segmentation. In order to exploit the information in the horizontal and vertical directions in images and remove redundant information to speed up, we propose the Asymmetric Pyramid Pooling Module (APPM) and the Asymmetric Factorized Module (AFM). TBANet with only 1M parameters processes 512×1024 images at 143 FPS and achieves 71.6% mIoU on the Cityscapes test set. TBANet stands out from other state-of-the-art networks in terms of accuracy and speed.