A Semi-Supervised Learning and Knowledge Distillation Based Method for Annotation of LIDAR Point Cloud Data

• Published in: 2023 4th International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE) pp. 14 - 22
• Main Authors: Yu, Chongchong, Liu, Fan, Chen, Jialun, Zheng, Tong, Feng, Wenbin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 25.08.2023
• Subjects: 3D guided filtering; 3D point cloud; Adaptation models; Filtering; knowledge distillation; Laser radar; Point cloud compression; point cloud semantic segmentation; Semantic segmentation; semi-supervised learning; Three-dimensional displays; Training
• DOI: 10.1109/ICBASE59196.2023.10303071

Lidar is a typical sensor in the fields of autonomous driving, environment-aware robots, and industrial automation. The 3D point cloud data collected by it contains rich 3D spatial information, and the surface description of objects is also close to reality, which helps people perceive complex scenes. In recent years, the deep learning-based lidar 3D point cloud data processing method has received widespread attention, realizing automatic feature learning from large amounts of data and realizing target detection and semantic segmentation of complex scenes. Among them, most methods need to learn model parameters based on a large amount of labeled data. However, the annotation of 3D point cloud data is more time-consuming and laborious than that of 2D images. Aiming at the above problems, this paper proposes a point cloud data labeling method based on the combination of semi-supervised learning and knowledge distillation. Firstly, the existing point cloud data is denoised by the improved three-dimensional guided filtering method to remove useless noise. Then, the classic Bilateral Augmentation and Adaptive Fusion (BAAF) point cloud semantic segmentation model is used as the teacher model. Moreover, to enhance the point cloud data, the local semantic context of the pyramid is introduced. We further generate high-quality pseudo-labels through the class range balance module. Finally, the rapid training of the student model is completed through the teacher averaging method in knowledge distillation and the obtained pseudo-labels. During the experiment, we discuss the labeling effect on the point cloud data of the public SemanticKITTI dataset and compare it with the scribble-annotated ScribbleKITTI data, to verify the ability of the proposed method.