Research map / Point clouds / Point cloud classification

MIRACLE 3D: Memory-efficient Integrated Robust Approach for Continual Learning on 3D Point Clouds via Shape Model Construction

ICLR 2025 · Point clouds › Point cloud classification · approach: Local feature aggregation networks

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 MIRACLE3D, a memory-efficient, privacy-preserving continual learning method for 3D point cloud classification that stores per-class shape models (mean shape plus a few SVD modes) to generate replay samples, with a Gradient Mode Regularization term.

Problem

Exemplar-based continual learning for 3D point cloud classification requires storing many raw exemplars, which is memory-heavy and raises privacy concerns, and herding-based exemplar selection is ill-suited to 3D feature spaces.

Approach

Represent each old class with a compact statistical shape model (mean shape plus top-k SVD modes of variation) and generate synthetic replay point clouds on demand by perturbing the mean along these modes with a small scale factor; add a regularizer penalizing the loss gradient projected onto the principal modes so the model is robust to shape variation.

Key results

• On ModelNet40, MIRACLE3D reaches 88.1% A_avg and 74.6% A_last with 120 stored samples versus 800 for I3DOL/InOR-Net and 120 for CL3D (69.7% A_last).
• On ShapeNet, MIRACLE3D reports 84.9% A_avg and 71.5% A_last; on ScanNet 72.9% A_avg and 57.2% A_last with 120 stored items.
• Gradient Mode Regularization improves ModelNet40 A_avg from 87.5% to 88.1% and A_last from 72.3% to 74.6%.

Measured effects reported in the paper

Surprising findings

• Storing 2 variation modes plus the mean (3 components per class) performs as well as or better than more modes, with 4 modes giving the same A_avg but a lower A_last than 2 modes in the ablation table.
• The reported BiC A_avg on ModelNet40 (48.5) is lower than its A_last (66.8), an inconsistent pattern relative to the other baselines.

Limitations

• Linear affine registration sometimes produces nonsensical or inaccurate shapes, especially for classes with high intra-class variability.
• Gains are modest relative to CL3D in some settings and depend on hyperparameters such as alpha=0.2 and the number of modes.
• Requires point clouds to be pre-registered and resampled into point correspondence, adding preprocessing (e.g., FilterReg for ScanNet).

Experimental setup

Datasets: ModelNet40 ShapeNet ScanNet

Benchmarks: ModelNet40 ShapeNet ScanNet

Metrics: average incremental accuracy (A_avg) last session accuracy (A_last) memory usage (M) forgetting rate (FR)

Baselines: LwF iCaRL DeeSIL EEIL IL2M DGMw DGMa BiC RPS-Net I3DOL InOR-Net CL3D

Builds on: iCaRL I3DOL CL3D Double Backpropagation Focal loss

Closest papers by meaning

• SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.86)
• Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · similarity 0.85)
• SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.85)
• Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.85)
• Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.84)
• GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.84)