Research map / Continual learning / Class-incremental
Activating Sparse Part Concepts for 3D Class Incremental Learning
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 ILPC, a 3D class-incremental learning framework that activates sparse shared part concepts, uses a part-relation Transformer head per task, and learnable affinities to fuse task heads.
Problem
Catastrophic forgetting in 3D class-incremental learning on point clouds, where classifiers rely on shortcut local features and task-head confusion causes old-class misclassification.
Approach
Learns a shared codebook of part concepts and activates only the top-K concepts per part feature to build part composition features; uses a per-task Transformer head over part features and learnable affinity terms to fuse task-wise classifiers instead of a fixed normalized sum.
Key results
- On Co3D, ILPC reaches 81.18 last and 86.37 avg accuracy, versus 78.87 last for the runner-up DS-AL.
- On nuScenes, ILPC reports 82.44 last / 87.27 avg, a gain of about 4 points last accuracy over the best prior method (DER 77.24 last, DS-AL 76.56).
- Ablation: adding affinity fusion with the Transformer head yields large gains (e.g., 18.8% and 22.5% in last accuracy across rows), and top-K=0.6 selection performs best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ILPC instead of DS-AL | PointNet-based 3D class-incremental learning · Co3D | last accuracy: 78.87 -> 81.18 | Helped |
| Use ILPC instead of DS-AL | PointNet-based 3D class-incremental learning · nuScenes | last accuracy: 76.56 -> 82.44 | Helped |
| Use ILPC instead of DS-AL | PointNet-based 3D class-incremental learning · ShapeNet | last accuracy: 80.96 -> 82.27 | Helped |
| Add a part codebook for part concept learning (ablation) | Baseline 3D CIL model · Co3D | last accuracy: 53.35 -> 58.42 | Helped |
| Replace mean pooling with the Transformer classification head (ablation) | Baseline 3D CIL model · Co3D | last accuracy: 53.35 -> 60.85 | Helped |
| Use affinity-based fusion instead of Eq. 5 fusion (ablation) | Baseline 3D CIL model · Co3D | last accuracy: 53.35 -> 56.99 | Helped |
| Add a part codebook to the Transformer-and-affinity model (ablation) | Transformer + affinity fusion · Co3D | last accuracy: 75.81 -> 78.93 | Helped |
| Add contrastive loss Lc (ablation) | Transformer + affinity fusion + part codebook · Co3D | last accuracy: 78.93 -> 79.97 | Helped |
| Add pseudo-part loss Lpp (ablation) | ILPC with Transformer, affinity fusion, and part codebook · Co3D | last accuracy: 79.97 -> 81.18 | Helped |
| Set TopK to 0.6 rather than 0.2, 0.4, 0.8, or 1.0 (ablation) | ILPC · Co3D | last accuracy: 78.87; 80.04; 81.18; 79.32; 78.18 | Mixed |
Surprising findings
- Changing backbones (PointNet++, DGCNN, PointNeXt) barely changes final accuracy of baselines, suggesting better backbones alone do not mitigate forgetting.
Limitations
- Learned part concepts are not guaranteed to be human-interpretable parts since no explicit part segmentation is enforced.
- Only point-cloud classification evaluated; no pretrained-model adaptation explored.
Experimental setup
Datasets: ShapeNetCore Co3D nuScenes
Benchmarks: ShapeNetCore Co3D nuScenes
Metrics: last accuracy average accuracy (avg)
Baselines: ER iCaRL DER FOSTER MEMO DS-AL DGR
Builds on: DNIK PointNet iCaRL DER
Closest papers by meaning
- Decompose Novel into Known: Part Concept Learning For 3D Novel Class Discovery (NeurIPS 2023 · similarity 0.88)
- PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding (NeurIPS 2025 · similarity 0.87)
- Reducing Class-wise Confusion for Incremental Learning with Disentangled Manifolds (CVPR 2025 · similarity 0.85)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.85)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.85)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.