Random Walks in Self-supervised Learning for Triangular Meshes

2025-03-02

AI Summary

This work addresses the insufficient geometric awareness and weak discriminability in self-supervised representation learning for 3D triangular meshes. To this end, we propose a structure-preserving data augmentation method based on random walks—introduced for the first time into mesh self-supervised learning. Furthermore, we design a synergistic optimization framework combining contrastive learning and K-means clustering losses, which simultaneously reduces training variance and enhances inter-class feature separability. Evaluated via SVM-based linear classification on shape retrieval and classification benchmarks, the learned representations achieve significantly higher mAP than leading unsupervised baselines; downstream transfer performance also consistently surpasses state-of-the-art methods. Our core contributions are: (1) a geometrically aware random-walk augmentation strategy that preserves intrinsic mesh topology and geometry; and (2) a novel joint contrastive–clustering training paradigm for self-supervised mesh representation learning.

Technology Category

• Machine Learning: Unsupervised & Self-Supervised Learning
• Computer Vision: Representation Learning for Vision
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding

Abstract

This study addresses the challenge of self-supervised learning for 3D mesh analysis. It presents an new approach that uses random walks as a form of data augmentation to generate diverse representations of mesh surfaces. Furthermore, it employs a combination of contrastive and clustering losses. The contrastive learning framework maximizes similarity between augmented instances of the same mesh while minimizing similarity between different meshes. We integrate this with a clustering loss, enhancing class distinction across training epochs and mitigating training variance. Our model's effectiveness is evaluated using mean Average Precision (mAP) scores and a supervised SVM linear classifier on extracted features, demonstrating its potential for various downstream tasks such as object classification and shape retrieval.

Problem

• Self-supervised learning for 3D mesh analysis
• Random walks for data augmentation in mesh surfaces
• Combining contrastive and clustering losses for class distinction

Innovation

• Random walks for 3D mesh data augmentation
• Contrastive and clustering losses combined
• Evaluation via mAP and SVM classifier