Abstract:The proliferation of depth cameras and LiDAR sensors in actual industrial environments has fueled the pursuit of an effective and efficient 3D point cloud model that enables us to perceive and interact with the physical world. However, the intrinsic complexity of 3D semantic information poses significant challenges to model design, including spatial rotation invariance and irregular point cloud structure, which fundamentally impact the representation and behavior of 3D point cloud systems. Existing have either heavily relied on labeling information in a supervised learning setting or failed to effectively capture the inherent patterns of the 3D point clouds within a self-supervised learning framework, leading to poor performance in specific downstream tasks. To address these limitations, this paper introduces a self-supervised framework, D ual- C ross C ontrastive Neural N etwork (DCCN) for 3D point cloud self-supervised representation learning. DCCN leverages cross-view, cross-network, and domain-specific knowledge distillation to enhance the extraction of hidden features from point clouds and fully exploit the capabilities of the encoder. Our DCCN employs a pseudo-Siamese network consisting of an online network and a target network, facilitating knowledge interaction and distillation. The method extracts internal states from augmented 3D point cloud by learning cross-view relationships and optimizes model parameters through intra-modal cross-network learning. We incorporate a momentum-updating mechanism without shared weights in the Siamese network architecture to distill knowledge and enhance the role differentiation the online and target networks. Experimental results demonstrate that our approach outperforms a range of supervised and self-supervised learning methods across a series of downstream tasks consisting of four tasks in three representative datasets. Ablation studies validate the component-wise effectiveness of cross-view, cross-network, and moment-updating learning objectives in achieving superior point cloud representation. The overall findings establish our method, DCCN, as an effective solution for 3D point cloud representation learning in real-world applications.

Is Contrastive Distillation Enough for Learning Comprehensive 3D Representations?

MultiCAD: Contrastive Representation Learning for Multi-modal 3D Computer-Aided Design Models

A Generalization Theory of Cross-Modality Distillation with Contrastive Learning

Extending Multi-modal Contrastive Representations

Connecting Multi-modal Contrastive Representations

Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning

Multi-Trusted Cross-Modal Information Bottleneck for 3D Self-Supervised Representation Learning

Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining

GroupContrast: Semantic-aware Self-supervised Representation Learning for 3D Understanding

Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models

P4Contrast: Contrastive Learning with Pairs of Point-Pixel Pairs for RGB-D Scene Understanding

Medical Vision-Language Representation Learning with Cross-Modal Multi-Teacher Contrastive Distillation

Self-Supervised Image-to-Point Distillation via Semantically Tolerant Contrastive Loss

DCCN: A dual-cross contrastive neural network for 3D point cloud representation learning

3D Shape Contrastive Representation Learning with Adversarial Examples

SegContrast: 3D Point Cloud Feature Representation Learning Through Self-Supervised Segment Discrimination

I$^2$MD: 3D Action Representation Learning with Inter- and Intra-modal Mutual Distillation

Cross-Modal Self-Supervised Learning with Effective Contrastive Units for LiDAR Point Clouds

Model2Scene: Learning 3D Scene Representation via Contrastive Language-CAD Models Pre-training

Sculpting Holistic 3D Representation in Contrastive Language-Image-3D Pre-training

Multimodal Contrastive Training for Visual Representation Learning