Balanced residual distillation learning for 3D point cloud class-incremental semantic segmentation

Yuanzhi Su, Siyuan Chen, Yuan-Gen Wang

Abstract: Highlights•We propose a novel Balanced Residual Distillation Learning (BRDL) framework for class-incremental semantic segmentation on 3D point clouds.•Residual Distillation Learning (RDL) effectively transfers and refines knowledge from previous classes while preventing catastrophic forgetting.•Balanced Pseudo-label Learning (BPL) mitigates class bias by promoting balanced learning between old and new classes during incremental training.•Extensive experiments on S3DIS and ScanNet datasets demonstrate the robustness and scalability of BRDL in 3D segmentation tasks.