Abstract
MBGRLp introduces a self-supervised framework for 3D point clouds that leverages graph contrastive learning to bypass the need for expensive manual labels. By combining cross-view and cross-model consistency checks, the model learns robust geometric features directly from raw point data, achieving performance comparable to fully supervised methods while avoiding the computational overhead of voxelization.
Introduction
3D point clouds are essential for spatial AI applications such as autonomous driving and urban planning. However, training deep learning models on this data faces two primary challenges: label scarcity and computational overhead. Supervised learning requires massive amounts of human-labeled 3D data, which is costly to produce. Furthermore, many existing methods convert irregular point clouds into regular 3D grids (voxels), a memory-intensive process that often discards fine-grained geometric details. MBGRLp addresses these issues by operating directly on raw point clouds through a self-supervised learning (SSL) approach.
Methodology
The MBGRLp architecture is inspired by graph representation learning frameworks and utilizes a "Bootstrap" approach similar to BYOL (Bootstrap Your Own Latent). The pipeline consists of three main stages:
- Augmentation: The system generates two distinct "views" of the same point cloud using random rotations, jittering, scaling, and shifting to ensure the model learns orientation-invariant features.
- Dual-Encoder Architecture: The framework employs an Online Encoder ($f_\theta$), which updates in real-time, and a Target Encoder ($f_\epsilon$), which is a stable version updated via Exponential Moving Average (EMA). This prevents model collapse without requiring negative samples.
- Multi-Scale Contrastive Learning: The model optimizes two loss functions: Cross-Model Loss ($L_{cn}$), comparing the online prediction to the target representation, and Cross-View Loss ($L_{cv}$), ensuring consistency between different views of the same object.
Experiments
The model was pre-trained on ShapeNetCore, a dataset containing approximately 51,300 unique 3D models across 55 categories. For evaluation, the authors utilized ModelNet40 and ModelNet10 (12,311 and 4,899 shapes, respectively). Performance was measured using a linear evaluation protocol, where the pre-trained encoder is frozen and only a simple final classification layer is trained.
Results
MBGRLp demonstrates highly competitive performance against both unsupervised and supervised baselines. On ModelNet40, the model achieved 90.22% accuracy, significantly outperforming other unsupervised methods like 3D-GAN (83.3%) and approaching the performance of the fully supervised DGCNN (92.9%).
Key findings include:
- Data Efficiency: Using only 10% of labeled downstream data, the model maintained an accuracy of 88.17% on ModelNet40, highlighting the robustness of the learned features.
- Ablation Studies: Removing either the cross-view or cross-model components resulted in a performance drop (to 87.43% and 88.70% respectively), confirming that both consistency checks are vital for learning effective representations.
Conclusion
MBGRLp successfully applies graph-based contrastive learning to raw 3D point clouds. By utilizing a bootstrapping mechanism and dual consistency checks, it achieves state-of-the-art results in self-supervised 3D object classification. Future work may explore more advanced backbones like Point Transformers and scaling the framework to large-scale scene segmentation.
Summary updated