Abstract:

With the rapid advancement of 3D understanding and computer vision technologies, point cloud data have gained significant attention for their precise geometry and rich spatial information. In applications such as intelligent transportation, accurately classifying subtle differences in vehicle models is crucial, rendering fine-grained point cloud classification essential. However, existing methods focus heavily on task-specific networks that enhance classification by extracting local discriminative features, often neglecting the model’s generalization abilities. This resulted in decreased performance in diverse scenarios and unseen categories, especially in environments with noise, occlusions, or data distribution changes. To address these challenges, the balanced generalization specialization network (BGS-Net) was proposed as a two-stage framework that balanced generalization and specialization. In the first stage, BGS-Net employed mask distillation self-supervised learning with coupled masks to guide two student models in learning independent feature representations from a teacher model, thereby enhancing generalization. In the second stage, a balanced training strategy was implemented by freezing one encoder to preserve general features while fine-tuning the other encoder to extract locally discriminative features for task specialization. Experimental results demonstrated that BGS-Net significantly outperformed existing methods in fine-grained, meta-category, few-shot, and real-world classification tasks, thereby confirming its effectiveness in maintaining high generalization while achieving task specialization. This approach enhanced the applicability and robustness of point cloud classification in practical applications.

Key words: point cloud understanding, self-supervised learning, coupling mask, feature learning, model generalization