EACNet: Efficient Asymmetric Cross-Modal RGB-D Semantic Segmentation Network for Indoor Robotic Perception

Efficient RGB-D semantic segmentation plays a crucial role in mobile robot environment perception and can enhance subsequent tasks. Given the context of indoor mobile robots’ operational environments and deployment settings, achieving effective RGB-D semantic segmentation necessitates a network architecture that is lightweight for efficiency, capable of real-time processing, and maintains high accuracy. Some existing methods fail to consider the differences between RGB and depth features or don’t fully extract cross-modal complementary information, resulting in lower segmentation accuracy; some employ complex network structures that cannot be deployed on mobile platforms and do not meet real-time requirements. To address the existing deficiencies, we propose EACNet, which leverages cross-modal correlations and complementary cues to achieve better performance with lower cost. Experimental results show EACNet achieves 57.6% mIoU on the NYUv2 datasets and 52.6% mIoU on the SUNRGBD datasets, with only 37M parameters and 40G FLOPs, indicating that our method achieves a better balance among segmentation accuracy, computation amount, and parameters than the state-of-the-art methods.