Abstract
Traditional self-supervised 3D learning methods often suffer from "semantic conflict" by treating different parts of the same object as negative pairs. GroupContrast addresses this by grouping points into semantically meaningful regions without human labels. By performing contrastive learning at the group level, it achieves state-of-the-art performance across 3D segmentation and object detection tasks on benchmarks like ScanNet and S3DIS.
Introduction: The Semantic Conflict in 3D Learning
Most current self-supervised 3D learning methods rely on point discrimination. When comparing two "views" of a 3D room, these methods attempt to match the exact same point in both views (positive pair) while pushing all other points away as negative pairs.
This approach creates a significant semantic conflict. For instance, if a model is forced to treat a point on the left side of a table as a "negative" to a point on the right side of the same table, it fails to learn that they belong to the same object. This leads to high false-negative rates and fragmented representations that struggle to identify objects as cohesive units.
Methodology
GroupContrast introduces a teacher-student framework designed to learn semantic groupings and contrastive features simultaneously.
1. Segment Grouping
The model first partitions the 3D point cloud into small geometric segments using a standard graph-cut method. It then employs a set of learnable prototypes (cluster centers) to group these segments into larger, semantically meaningful regions. This process uses "informative-aware distillation," where a teacher network guides a student network to assign segments to the same prototypes across different views.
2. Semantic-aware Contrastive Learning
Once groups are formed, the model performs contrastive learning where points within the same group are treated as positive pairs, and points in different groups are negatives. To handle initial noise, a confidence weight is introduced, ensuring the model does not over-rely on uncertain groupings during early training phases.
3. Architecture
The system uses a SparseUNet backbone. The teacher network is updated via an Exponential Moving Average (EMA) of the student’s weights, providing a stable target for learning.
Experiments and Results
The researchers evaluated GroupContrast on major 3D indoor scene benchmarks, including ScanNet, ScanNet200, S3DIS, and SUN RGB-D.
Performance Benchmarks
GroupContrast established new state-of-the-art results for self-supervised 3D representation:
- Semantic Segmentation: Achieved 75.7% mIoU on ScanNet, outperforming previous methods and the "from-scratch" baseline by 3.5%.
- Instance Segmentation: Reached 27.5 [email protected] on ScanNet200.
- Object Detection: Improved 3D object detection on ScanNet to 41.1 [email protected], surpassing the previous best by 1.8 points.
Key Insights
- Emergent Concepts: Learnable prototypes naturally begin to represent real-world concepts like "floors," "chair backrests," or "ceilings" without any human labels.
- Data Efficiency: In low-label regimes (1% or 5% labeled data), GroupContrast significantly outperforms traditional methods, demonstrating high efficiency.
- Foundation Model Synergy: The grouping method can be enhanced by integrating masks from 2D foundation models like Segment Anything (SAM), further boosting performance to 75.9% mIoU.
Conclusion
GroupContrast demonstrates that seeking semantic coherence is more effective than simple geometric point discrimination for 3D understanding. By moving beyond point-to-point comparisons, the framework provides a more robust foundation for complex 3D tasks. While currently focused on indoor scenes, future work aims to scale this approach to more diverse datasets and universal 3D foundation models.
Summary updated