Semantic Completion and Rendering of 3D Scenes to Improve Downstream Model Performance

This thesis addresses the challenge of enhancing downstream model performance in semantic segmentation tasks by tackling the prevalent issue of incomplete or sparse 3D data in real-world scenarios. Our work focuses on semantic completion and rendering of 3D scenes as a crucial preceding data augmentation step. The approach utilizes advanced machine learning techniques, particularly generative models and neural rendering, to synthesize missing scene information, thereby improving the data's representation. By providing enriched 3D representations, this methodology aims to increase the accuracy of downstream tasks, including object detection. We demonstrate the effectiveness of this strategy by showcasing various performance metrics and its potential to bridge the gap between ideal synthetic environments and noisy real-world sensor data.