3D diffusion based generation model for point cloud annotation and generation.

Abstract

Concurrent with the rapid advancement of applications and 3D scanning sensors, the demand for 3D deep learning based technology and data has increased dramatically. Especially 3D shape with semantic labels plays a significant role in 3D vision problems, such as auto-driven, 3D object detection and 3D scene segmentation, etc. As the deep learning era arrives, automatic, high-quality, and large-scale solutions in annota- tion 3D shape to the 3D vision problem are desired. Point cloud, as one of the most popular representations of 3D, is facing the same desire. Point cloud generative model is one type of model that can be used to synthesize a new point cloud. The characteristic of the point cloud generative model indicates that it contains the semantic structure of a point cloud. The interrelationships of point cloud attract many researchers to explore to use of point cloud generative to solve annotated point cloud acquire problems. However, it is still challenging to acquire expressive and accurate annotated point clouds. This thesis addresses the aforementioned challenge by explor- ing three aspects: the synthesis of high-quality 3D point cloud objects, the point-label pairs generation, and the evolution of their enhancement strategies. • This work introduces a point cloud diffusion generation model combining stochastic differential equations and Markov Chain Mento Carlo samplers. This method can synthesise high-quality 3D point cloud objects and al- lows a more flexible sampling method to point cloud generation. • Furthermore, the thesis presents a point-label pairs gen- eration method to alleviate the cost of large-scale point cloud annotation. This method investigates the charac- teristics of diffusion-based point cloud generation model and exploits a feature interpreter to generate a point cloud with corresponding semantic labels for each point. • Last, a filter approach for generated point-label pairs is employed to improve the quality of the generated point cloud dataset. As a result, the proposed method resolves the point cloud generation and annotation effectively. To demonstrate the effectiveness of the proposed method, various experiments were conducted across different scenar- ios. These experiments not only validated the reliability of the generated point cloud and point-label pairs but also il- lustrated their superior performance in comparison to GAN- based point-label generation methods. This research repre- sents a substantial contribution to the enhancement of the quality and applicability of 3D point cloud data and under- standing.