Research map / Point clouds / Point cloud classification
LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling
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 LCM, a compact point cloud backbone (locally constrained kNN aggregation encoder plus Mamba decoder with a local-constraint FFN) that replaces Transformers in masked point modeling, cutting parameters and FLOPs while improving ScanObjectNN classification.
Problem
Transformer-based masked point modeling for point clouds has quadratic attention complexity and large parameter counts, hindering deployment on embedded devices, and its decoder handles the mismatch in information density between unmasked and masked tokens poorly.
Approach
Replace self-attention in the encoder with a local aggregation layer using static kNN geometric neighborhoods (two MLPs, computed once and shared across layers), and replace the decoder's self-attention with a Mamba SSM followed by a Local Constraints Feedforward Network that implicitly exchanges information among geometrically adjacent patches, removing dependence on input sequence order.
Key results
- Point-MAE w/ LCM reaches 94.51% on ScanObjectNN OBJ-BG, +1.84 over the Transformer-based Point-MAE, with 2.7M params vs 22.1M and 1.3G FLOPs vs 4.8G.
- On ScanNetV2 detection, Point-MAE-based LCM improves AP25 by 5.2% and AP50 by 6.0% over the Transformer encoder (as stated in the paper).
- Trained from scratch, LCM improves ScanObjectNN classification by 0.82, 0.15, and 1.10 points over the Transformer with 88% fewer parameters and 73% fewer FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the Transformer with LCM in Point-MAE | Point-MAE · ScanObjectNN OBJ-BG, OBJ-ONLY, and PB-T50-RS | accuracy: 92.67 -> 94.51; 92.08 -> 92.75; 88.27 -> 88.87 | Helped |
| replace the Transformer encoder with the compact LCM encoder | 3DETR with Point-MAE pretraining · ScanNetV2 | AP25: +5.2% | Helped |
| replace the Transformer encoder with the compact LCM encoder | 3DETR with Point-MAE pretraining · ScanNetV2 | AP50: +6.0% | Helped |
| add local aggregation to an encoder with a simple two-layer MLP (ablation) | LCM compact encoder · ScanObjectNN | classification performance | Helped |
| add a simple two-layer MLP without local aggregation (ablation) | LCM compact encoder · ScanObjectNN | classification performance | No clear effect |
| add FFN to the local-aggregation encoder (ablation) | LCM compact encoder · ScanObjectNN | classification performance | Helped |
| increase the number of local neighbors K (ablation) | LCM compact encoder · ScanObjectNN PB-RS-T50 | classification performance | Mixed |
| use feature-space K-NN instead of geometric-space K-NN (ablation) | LCM compact encoder · ScanObjectNN | classification performance | Hurt |
| use different patch sequence orders (ablation) | Mamba encoder · ScanObjectNN PB-RS-T50 | accuracy: 82.34% vs. 80.65%; difference of 1.69% | Mixed |
| add LCFFN to the Mamba decoder (ablation) | Mamba decoder · ScanObjectNN PB-RS-T50 | classification performance | Helped |
| replace the vanilla Mamba decoder's FFN with LCFFN (ablation) | Mamba decoder · ScanObjectNN PB-RS-T50 | runtime | Helped |
| use combined patch sequences instead of a single sequence order (ablation) | Mamba encoder · ScanObjectNN PB-RS-T50 | classification performance | Helped |
Surprising findings
- Vanilla Mamba decoder is sensitive to input patch ordering, with a 1.69% accuracy gap between orders; adding LCFFN removes most of this dependence.
- Reproducing prior pretraining methods under the authors' downstream setting raised their ScanObjectNN accuracy substantially (e.g., Point-BERT up 5.34% on OBJ-BG), suggesting earlier baselines were under-tuned.
Limitations
- Does not model dynamic importance or long-range dependencies, as acknowledged by the authors.
- Naive top-K attention still has quadratic complexity; the efficiency gain depends on the redesigned local layers.
- Improvements on ScanObjectNN are modest (0.60-1.84 points) and depend on the authors' re-tuned downstream setting.
Experimental setup
Datasets: ShapeNet ScanObjectNN ModelNet40 ScanNetV2 ShapeNetPart
Benchmarks: ScanObjectNN ModelNet40 ScanNetV2 ShapeNetPart
Metrics: overall accuracy average accuracy AP25 AP50 FLOPs parameters
Baselines: Transformer Point-BERT MaskPoint Point-MAE Point-M2AE ACT PointMamba PointNet++ PointGPT-B
Builds on: Point-MAE Mamba Transformer PointMamba 3DETR
Closest papers by meaning
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.90)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.89)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.88)
- Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.87)
- RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.87)
- General Point Model Pretraining with Autoencoding and Autoregressive (CVPR 2024 · similarity 0.87)
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.