Semi-övervakad klassificering av 3D-punktmolns form

The widespread integration of Lidar scanners and software solutions within autonomous vehicle technology has highlighted the necessity for enhanced comprehension of the environment and detection of objects. Despite the notable progress in 2D object classification using deep learning, the development of learning-based 3D object classification has been relatively sluggish due to the inherent difficulty in creating precise annotations for three-dimensional objects. To overcome this challenge, there is a growing interest in harnessing semi-supervised learning methods in deep learning, which capitalize on both labeled and unlabeled data for training, offering a cost-effective and time-saving alternative to traditional labeling approaches. This thesis explores the fusion of the FixMatch semi-supervised learning framework [1] with Pointnet++ [2] for 3D point cloud shape classification. The classifier undergoes training on two distinct datasets – ModelNet40 representing synthetic data and Flasheye objects derived from real-world scenarios. A comparative analysis of the models trained on both synthetic and real-world datasets is conducted against a supervised baseline. The findings indicate that the semi-supervised learning framework enhances model performance, even in scenarios with limited labeled data.