Research map / Point clouds / Point cloud classification

PMA: Towards Parameter-Efficient Point Cloud Understanding via Point Mamba Adapter

CVPR 2025 · Point clouds › Point cloud classification · approach: Self-supervised point cloud pretraining · 7 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 Point Mamba Adapter (PMA), a parameter-efficient fine-tuning method that fuses intermediate-layer features of frozen point cloud Transformers via a Mamba adapter whose output gate is modulated by a shared geometry-constrained gate prompt generator (G2PG).

Problem

Parameter-efficient fine-tuning of pre-trained point cloud models typically uses only the final-layer features, discarding complementary intermediate-layer information, which hurts fine-grained tasks like part segmentation.

Approach

Build an ordered sequence of tokens from all layers of a frozen point Transformer and fuse them with a Mamba adapter; a shared G2PG produces geometric gate prompts for Mamba's output matrix C and dynamically learns a token ordering via KNN-based soft assignment and argmax, addressing the isotropy of 3D space.

Key results

• On ScanObjectNN PB-T50-RS, PointGPT-L w/ PMA reaches 94.9% (vs 94.7 FFT) with 4.9M trainable parameters, a ~99% reduction from full fine-tuning.
• Point-MAE w/ PMA achieves 86.43% on PB-T50-RS, above IDPT (84.94), DAPT (85.08), and FFT (85.18) with 1.1M trainable parameters.
• Ablation on PB-T50-RS: adding the Mamba adapter raises accuracy from 76.72% to 85.08%, with gate prompt and dynamic reorder reaching 86.43%.

Measured effects reported in the paper

Surprising findings

• Using only the first few layers of the frozen Point-MAE backbone yields performance nearly identical to using all layers (drops of 0.14% for segmentation with 3 layers).

Limitations

• Gains on some benchmarks are small (e.g., +0.13 on OBJ-ONLY for PointGPT-L) and PMA does not win on every variant (e.g., PB-T50-RS in Point-BERT results is comparable to baselines).
• Reported few-shot and segmentation improvements are modest relative to standard deviations.
• Evaluated only on a limited set of point cloud pre-trained backbones and tasks.

Experimental setup

Datasets: ScanObjectNN ModelNet40 ShapeNetPart

Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart

Metrics: overall accuracy mIoU few-shot accuracy

Baselines: Full fine-tuning IDPT DAPT PointGST Adapter LoRA VPT AdaptFormer BitFit Point-BERT Point-MAE Recon

Builds on: Mamba PointGPT Point-MAE IDPT DAPT

Closest papers by meaning

• Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.91)
• GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.89)
• Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.88)
• PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.88)
• Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception (CVPR 2025 · similarity 0.87)
• Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.87)