Unsupervised Training Framework for 3D Point Cloud Object Detection Model
Wu, Min-Jyun Lin, Hsiang-Jui Hung, Shih-Hao Shih, Chi-Sheng
Wu, Min-Jyun
Lin, Hsiang-Jui
Hung, Shih-Hao
Shih, Chi-Sheng
Supervisor
Department
Computer Science
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Point cloud data frames are critical, if not indispensable, for precise robot navigation and localization, but training the object detection models for them remains challenging. Many models require labeled 3D objects to train the model. However, the sparse and occluded 3D point cloud data make it difficult, if not impossible, to automate the labeling process. This work proposes a training framework to generate 3D labels on point clouds to tackle the aforementioned challenges. The proposed method takes advantage of the consecutive presence of the same object on different frames to automate the labeling process. The experimental results show that the unsupervised framework trains a robust model for 3D object detection. On the roadside data, the model archives 90.27% AP for scooters and 91.33% AP for cars. On nuScenes dataset, the framework demonstrates the detection precision doubles on IoU 0.25 and IoU 0.5 when the recalls remain similar, compared to the model trained by the MODEST framework.
Citation
M.-J. Wu, H.-J. Lin, S.-H. Hung, and C.-S. Shih, “Unsupervised Training Framework for 3D Point Cloud Object Detection Model,” Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing, pp. 1146–1155, Mar. 2025, doi: 10.1145/3672608.3707799.
Source
Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing
Conference
SAC '25: 40th ACM/SIGAPP Symposium on Applied Computing
Keywords
Computer Systems Organization, Computer Communication Networks
Subjects
Source
SAC '25: 40th ACM/SIGAPP Symposium on Applied Computing
Publisher
Association for Computing Machinery