We present PointContrast, a framework for unsupervised pre-training of 3D point cloud representations. We show that contrastive learning can be effectively applied to 3D point clouds by leveraging correspondences between different views of the same scene. Points that correspond to the same 3D location across views serve as positive pairs, while non-corresponding points serve as negative pairs. Pre-training with PointContrast significantly improves downstream 3D tasks including detection and segmentation on ScanNet and S3DIS.

Self-supervised learning has achieved remarkable success in 2D vision and NLP, but its application to 3D point clouds remains largely unexplored. Unlike images, point clouds are unordered, sparse, and vary in density, making it unclear how to define effective pretext tasks. We propose to exploit the natural multi-view structure of 3D scenes: the same physical point observed from different viewpoints provides free supervision for contrastive learning.

The challenge of applying contrastive learning to 3D data lies in defining meaningful positive and negative pairs. In 2D, augmentations of the same image form positive pairs. For 3D point clouds, we leverage geometric correspondences from multi-view registration: given two partial scans of the same scene, points in the overlapping region that map to the same 3D coordinate are natural positive pairs. This formulation is geometrically grounded and does not require synthetic augmentations.

Given a 3D scene, we create two partially overlapping point cloud views by random spatial transformations. Points in the overlap region that correspond to the same 3D location are positive pairs. We extract features using a sparse convolutional network (e.g., SparseConvNet) and apply a PointInfoNCE contrastive loss that pulls positive pairs together and pushes negative pairs apart in the feature space. The pre-trained backbone is then fine-tuned on downstream tasks.

The PointInfoNCE loss is formulated as: for each positive pair (i, j), maximize the similarity of their features relative to all negative pairs. We use hardest-contrastive sampling to focus on the most informative negatives. We also study the effect of pre-training data scale and find that performance improves log-linearly with the amount of pre-training data, suggesting that more unlabeled 3D data can further improve results.

On ScanNet 3D detection, PointContrast pre-training improves VoteNet by +3.4 mAP@0.5. On ScanNet semantic segmentation, it improves by +2.7 mIoU. On S3DIS segmentation, improvement is +2.2 mIoU. These gains are consistent across different architectures (SparseConvNet, MinkowskiNet), demonstrating the generality of the pre-trained representations. PointContrast also outperforms other 3D pre-training methods including autoencoder-based approaches.

We further analyze the data efficiency of PointContrast pre-training. When fine-tuning with only 20% of labeled data, the pre-trained model achieves performance comparable to training from scratch with 100% labeled data. This demonstrates that PointContrast learns highly transferable representations that significantly reduce the annotation requirement for 3D understanding tasks.

We have demonstrated that contrastive pre-training is highly effective for 3D point cloud understanding. PointContrast leverages the natural multi-view structure of 3D data to learn transferable representations. Our results show that the paradigm of self-supervised pre-training followed by supervised fine-tuning, successful in 2D and NLP, translates effectively to the 3D domain.