ADBNet: Asymmetric dual-branch network for indoor real-time RGB-D semantic segmentation

Real-time RGB-D semantic segmentation is critical for a wide range of visual scene understanding tasks. Although recent methods have achieved notable progress, they still face significant challenges in fully exploiting depth information and achieving robust RGB-D fusion, particularly under the constraints of real-time performance. Balancing segmentation accuracy and computational efficiency remains a key bottleneck for practical deployment. To address these challenges, we propose ADBNet, an efficient asymmetric dual-branch network for real-time indoor semantic segmentation. ADBNet employs an asymmetric encoder with enhanced depth processing and incorporates a novel Conv-Former Pyramid Vision Transformer (CF-PVT) featuring decomposed convolutional attention to improve depth feature extraction. Furthermore, an Adaptive Feature Recalibration and Fusion (AFRF) module is introduced to enable effective cross-modal alignment and multi-scale feature fusion. Experiments on the NYU-Depth V2 dataset demonstrate that ADBNet achieves an excellent trade-off between accuracy and efficiency, running at 66 FPS with 79.40% pixel accuracy and 56.0% mean IoU.