Exploring 3D Dataset Pruning

Dataset pruning has been widely studied for 2D images to remove redundancy and accelerate training, while particular pruning methods for 3D data remain largely unexplored. In this work, we study dataset pruning for 3D data, where its observed common long-tail class distribution nature make optimization under conventional evaluation metrics Overall Accuracy (OA) and Mean Accuracy (mAcc) inherently conflicting, and further make pruning particularly challenging. To address this, we formulate pruning as approximating the full-data expected risk with a weighted subset, which reveals two key errors: coverage error from insufficient representativeness and prior-mismatch bias from inconsistency between subset-induced class weights and target metrics. We propose representation-aware subset selection with per-class retention quotas for long-tail coverage, and prior-invariant teacher supervision using calibrated soft labels and embedding-geometry distillation. The retention quota also serves as a switch to control the OA-mAcc trade-off. Extensive experiments on 3D datasets show that our method can improve both metrics across multiple settings while adapting to different downstream preferences. Our code is available at https://github.com/XiaohanZhao123/3D-Dataset-Pruning.

• Publication year

  2026

• Venue

  Unknown venue

• Publication date

  2026-02-28

• Fields of study

  Computer Science

• Identifiers
  arXiv 2603.00651
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• Medium-Difficulty Samples Constitute Smoothed Decision Boundary for Knowledge Distillation on Pruned Datasets

  2025cited by this paper
• Class-Proportional Coreset Selection for Difficulty-Separable Data

  2025cited by this paper
• Non-Uniform Class-Wise Coreset Selection: Characterizing Category Difficulty for Data-Efficient Transfer Learning

  2025influential reference
• Distilling the Knowledge in Data Pruning

  2024cited by this paper
• DRoP: Distributionally Robust Data Pruning

  2024influential reference
• Masked Autoencoders for 3D Point Cloud Self-supervised Learning

  2023cited by this paper
• Bilevel Coreset Selection in Continual Learning: A New Formulation and Algorithm

  2023cited by this paper
• Squeeze, Recover and Relabel: Dataset Condensation at ImageNet Scale From A New Perspective

  2023cited by this paper
• MeshMAE: Masked Autoencoders for 3D Mesh Data Analysis

  2022cited by this paper
• Scaling Up Dataset Distillation to ImageNet-1K with Constant Memory

  2022cited by this paper
• PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

  2022cited by this paper
• Dataset Pruning: Reducing Training Data by Examining Generalization Influence

  2022cited by this paper
• GRAD-MATCH: Gradient Matching based Data Subset Selection for Efficient Deep Model Training

  2021cited by this paper
• Foundations of Machine Learning

  2021influential reference
• Deep Learning on a Data Diet: Finding Important Examples Early in Training

  2021cited by this paper
• Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World Data

  2019cited by this paper
• Class-Balanced Loss Based on Effective Number of Samples

  2019cited by this paper
• Variational Adversarial Active Learning

  2019cited by this paper
• Relational Knowledge Distillation

  2019influential reference
• Decoupling Representation and Classifier for Long-Tailed Recognition

  2019cited by this paper
• Coresets for Data-efficient Training of Machine Learning Models

  2019cited by this paper
• MeshNet: Mesh Neural Network for 3D Shape Representation

  2018cited by this paper
• An Empirical Study of Example Forgetting during Deep Neural Network Learning

  2018cited by this paper
• Active Learning for Convolutional Neural Networks: A Core-Set Approach

  2017influential reference
• PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

  2017cited by this paper
• A systematic study of the class imbalance problem in convolutional neural networks

  2017cited by this paper
• Spectrally-normalized margin bounds for neural networks

  2017cited by this paper
• A scalable active framework for region annotation in 3D shape collections

  2016influential reference
• Distilling the Knowledge in a Neural Network

  2015cited by this paper
• ShapeNet: An Information-Rich 3D Model Repository

  2015cited by this paper
• Submodularity in Data Subset Selection and Active Learning

  2015influential reference
• 3D ShapeNets: A deep representation for volumetric shapes

  2014cited by this paper
• A new active labeling method for deep learning

  2014cited by this paper
• Herding dynamical weights to learn

  2009cited by this paper
• Integral Probability Metrics and Their Generating Classes of Functions

  1997influential reference

• No citing papers are available for this paper.