APPFNet: Adaptive point-pixel fusion network for 3D semantic segmentation with neighbor feature aggregation

3D semantic segmentation is significant for scene understanding in various domains, such as autonomous driving, mapping, and robotics. Existing research often enhances prediction accuracy by integrating data from camera and LiDAR (light detection and ranging). However, current fusion methods face two primary challenges. Firstly, they frequently project point clouds onto 2D coordinates and employ 2D segmentation networks, resulting in a significant loss of spatial depth information. Secondly, prevailing approaches struggle to establish long-range contextual understanding, limiting comprehensive predictions across multiple objects. To address these challenges, our work proposes the Adaptive Point-Pixel Fusion Network (APPFNet). In the initial stage of modality fusion, this method correlates points with pixels to provide spatial depth information to Red, Green, and Blue (RGB) images. Subsequently, we leverage our adaptive Multi-scale Modal Fusion module (MSMFM) and Dispersed Self-attention (DFSA) mechanism for further feature extraction and fusion. Additionally, we propose a flexible Neighbor Feature Aggregation module (NFAM) that enhances the model’s ability to establish long-range contextual understanding by fusing information from neighboring points. Importantly, this module can be easily transplanted into other networks with simple modifications. Extensive experiments on SemanticKITTI and nuScenes demonstrate the superiority of our proposed approach, where APPFNet achieves a higher mean accuracy (mAcc) in predicting small objects in distant areas compared to state-of-the-art (SOTA) fusion methods. •A multi-modal network architecture APPFNet is proposed based on Transformer.•A new module NFAM is proposed to enhance point cloud information.•APPFNet performs semantic segmentation using images and complete point clouds.•The model exhibits higher accuracy for distant objects compared to existing methods.