Dual-level Adaptive Self-Labeling for Novel Class Discovery in Point Cloud Segmentation
This work solves the problem of discovering novel classes in point cloud segmentation for applications like autonomous driving, but it is incremental as it builds on existing self-labeling and representation techniques.
The paper tackles novel class discovery in point cloud segmentation by addressing imbalanced class distributions and spatial context neglect, resulting in a method that outperforms state-of-the-art approaches on SemanticKITTI and SemanticPOSS datasets.
We tackle the novel class discovery in point cloud segmentation, which discovers novel classes based on the semantic knowledge of seen classes. Existing work proposes an online point-wise clustering method with a simplified equal class-size constraint on the novel classes to avoid degenerate solutions. However, the inherent imbalanced distribution of novel classes in point clouds typically violates the equal class-size constraint. Moreover, point-wise clustering ignores the rich spatial context information of objects, which results in less expressive representation for semantic segmentation. To address the above challenges, we propose a novel self-labeling strategy that adaptively generates high-quality pseudo-labels for imbalanced classes during model training. In addition, we develop a dual-level representation that incorporates regional consistency into the point-level classifier learning, reducing noise in generated segmentation. Finally, we conduct extensive experiments on two widely used datasets, SemanticKITTI and SemanticPOSS, and the results show our method outperforms the state of the art by a large margin.