Research map / Point clouds / Point cloud classification

MoST: Efficient Monarch Sparse Tuning for 3D Representation Learning

CVPR 2025 · Point clouds › Point cloud classification · approach: Point-language and open-vocabulary 3D · 0 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 MoST, a reparameterization-based PEFT method for 3D point cloud backbones that replaces dense update matrices with sparse, locally-aware Point Monarch matrices, adding no inference overhead.

Problem

Parameter-efficient fine-tuning of pretrained 3D point cloud models; existing 3D PEFT methods rely on adapters or prompts that add inference overhead and generalize poorly beyond Transformers, while LoRA-style reparameterization misses local geometric features.

Approach

Extends Monarch structured matrices to unstructured points with K-Rectify, a KNN-based local token linear transformation applied before and after the block-diagonal Monarch factors, yielding Point Monarch matrices that reparameterize dense update weights and can be merged at inference.

Key results

• MoST on PointGPT reaches 97.50% accuracy on ScanObjectNN PB 50 RS and 96.23% on ModelNet40 with 0.8M (b=32) to 8.0M trainable parameters for b=8.
• With only 0.8M (3.6%) trainable parameters, MoST (b=32) on Point-MAE reaches 91.95% on ScanObjectNN, surpassing full fine-tuning (85.18%) by 6.77%.
• MoST on ReCon gives 92.85% ScanObjectNN accuracy with 1.3M parameters, outperforming the PEFT baselines IDPT, DAPT, PPT and PointGST.

Measured effects reported in the paper

Surprising findings

• LoRA (82.76%) underperforms full fine-tuning and several 3D PEFT adapters on ScanObjectNN, and MoST's gains over LoRA are attributed to the local K-Rectify rather than the reparameterization alone.

Limitations

• Gains over full fine-tuning are reported on specific backbones; some PEFT results on other backbones (e.g., Transformer sparse training on S3DIS) trail full fine-tuning.
• Block count b and K-Rectify hyperparameters require tuning; performance peaks at b=8 and degrades at larger b.

Experimental setup

Datasets: ScanObjectNN PB 50 RS ModelNet40 ShapeNetPart S3DIS Area 5 ShapeNet

Benchmarks: ScanObjectNN PB 50 RS ModelNet40 ShapeNetPart S3DIS Area 5 ModelNet40 few-shot

Metrics: overall accuracy class mIoU instance mIoU mAcc trainable parameters FLOPs

Baselines: LoRA BitFit IDPT DAPT PPT PointGST VPT Full fine-tuning

Builds on: Monarch matrices LoRA Kronecker decomposition Butterfly matrices

Closest papers by meaning

• Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models (AAAI 2024 · similarity 0.87)
• On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation (NeurIPS 2025 · similarity 0.85)
• Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · similarity 0.85)
• Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.84)
• SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.84)
• LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.84)