Research map / Point clouds / Point cloud classification

PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning

CVPR 2025 · Point clouds › Point cloud classification · approach: Self-supervised point cloud pretraining · 3 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 PointLoRA, which combines LoRA injected into the qkv and FFN layers of point cloud transformers with a multi-scale top-K token selection module that provides local geometric prompts, reaching competitive results with 3.43% trainable parameters.

Problem

Parameter-efficient fine-tuning of pre-trained point cloud transformers, where existing PEFT methods rely on complex adapters and prompts that add tunable parameters and compute.

Approach

Inject low-rank adaptation into the parameter-heavy qkv and FFN layers to capture global features, and complement it with multi-scale FPS/kNN tokens selected by a learned Mask Predictor and fed through a shared Prompt MLP as local geometry prompts.

Key results

• On PB-T50-RS, PointLoRA with Point-MAE reaches 85.53% OA using 0.77M tunable parameters (3.43%), surpassing full fine-tuning (85.18%) and DAPT (85.08%).
• With rank 8, PointLoRA reaches 85.53% on PB-T50-RS; token selection adds about 1 point over LoRA alone (83.83% to 84.91%, before multi-scale FPS).
• On ShapeNetPart with ReCon, PointLoRA reaches 83.98% class mIoU and 85.4% instance mIoU with 5.63M tunable parameters.

Measured effects reported in the paper

Surprising findings

• Generic NLP/2D PEFT methods such as Adapter, LoRA, and VPT give limited gains on point clouds, with LoRA alone reaching 81.74% on PB-T50-RS.

Limitations

• Gains over prior PEFT methods are small on several benchmarks (e.g., ModelNet40 few-shot and ShapeNetPart instance mIoU match or trail baselines).
• Hyperparameters such as rank, token counts and scale number were tuned on the PB-T50-RS ablation split.

Experimental setup

Datasets: ScanObjectNN ModelNet40 ShapeNetPart

Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart

Metrics: overall accuracy mIoU tunable parameters

Baselines: Full fine-tuning Linear probing IDPT DAPT PPT Point-PEFT Adapter VPT-Deep AdaptFormer SSF

Builds on: LoRA Point-MAE ReCon VPT IDPT

Closest papers by meaning

• Dynamic Adapter Meets Prompt Tuning: Parameter-Efficient Transfer Learning for Point Cloud Analysis (CVPR 2024 · similarity 0.88)
• $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.88)
• Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models (AAAI 2024 · similarity 0.86)
• GeoLoRA: Geometric integration for parameter efficient fine-tuning (ICLR 2025 · similarity 0.86)
• RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
• Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.85)