Treffer: 3DOS Error Analysis and 3D Ultrasound Scanning Application Proposal

The widespread application of 3D point cloud analysis has garnered significant research interest, particularly regarding the handling of out-of-distribution (OOD) data, which is crucial for real-world safety-demanding applications. To enhance neural network robustness, this report investigates mechanisms based on error analysis. Specifically, we reproduce key results from 3DOS: Towards 3D Open Set Learning -- Benchmarking and Understanding Semantic Novelty Detection on Point Clouds and conduct a statistical analysis to understand the underlying causes of underconfidence and overconfidence, which contribute to performance degradation. We compare various methods across different families and backbones. Additionally, we propose a real-world application for 3D Semantic Novelty Detection.