Research map / Continual learning / Class-incremental

LAGD: Local Topological-Alignment and Global Semantic-Deconstruction for Incremental 3D Semantic Segmentation

AAAI 2025 · Continual learning › Class-incremental · approach: Unknown-class mining for incremental segmentation · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LAGD, a class-incremental 3D semantic segmentation framework combining interaction topological-aware alignment of local geometric structures with semantic-knowledge deconstruction distillation that splits logits into positive and negative parts.

Problem

Catastrophic forgetting in class-incremental 3D point cloud semantic segmentation, where new categories arrive continually and existing 2D CISS methods aggravate forgetting due to irregular geometry and global semantic (logit) shift.

Approach

Models state-specific knowledge via cross-attention exchange between old and new encoder features, aligns k-NN-based local topological structures of those features (ITA), and distills positive and negative logit parts separately (SDD) instead of coupled vanilla KD.

Key results

• On ScanNet 10-1 long-term setting (11 steps), LAGD reaches 43.7% average mIoU vs 36.9% for LGKD and 36.3% for F&A, and 8.5% forgetting-drop vs 44.2% for fine-tuning.
• On S3DIS Cnovel=5 M0 split, LAGD achieves 47.8% overall mIoU, above LGKD (21.8%) and LwF (43.5%) reproductions.
• Ablation on S3DIS (M0) shows adding ITA and SDD on top of Lkd raises overall mIoU from 43.5 to 47.8.

Measured effects reported in the paper

Surprising findings

• Naively applying 2D class-incremental methods to 3D point clouds can aggravate forgetting, as reported for typical CIL approaches.
• On the ScanNet short-term M1 setting, LGKD performs poorly on several splits, while LAGD's gains are largest in the Cnovel=5 and long-term settings.

Limitations

• ITA uses features only from the first encoder layer for efficiency, which may limit the topological information captured.
• Performance drops notably for Cnovel=1 settings and under the M1 split on S3DIS compared with M0, showing sensitivity to class order.
• Hyperparameters alpha, beta, gamma are fixed at {10,1,1} without reported sensitivity analysis.

Experimental setup

Datasets: S3DIS ScanNet

Benchmarks: S3DIS ScanNet

Metrics: mIoU

Baselines: Fine-Tune Joint-training Freeze and Add LwF EWC LGKD GUAT

Builds on: LwF LGKD GUAT PointNet++

Closest papers by meaning

• BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis (CVPR 2025 · similarity 0.87)
• UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.86)
• Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation (NeurIPS 2026 · similarity 0.86)
• Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning (NeurIPS 2026 · similarity 0.86)
• SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.86)
• econSG: Efficient and Multi-view Consistent Open-Vocabulary 3D Semantic Gaussians (ICLR 2025 · similarity 0.86)