PreSem-Surf introduces an optimized approach for surface reconstruction using RGB-D sequences by leveraging the Neural Radiance Field (NeRF) framework. The method uniquely combines RGB, depth, and semantic information to enhance reconstruction quality. A key innovation is the SG-MLP sampling structure paired with a Preconditioning Multilayer Perceptron (PR-MLP) for voxel pre-rendering, which enables earlier capture of scene-related features and better differentiation between noise and local details. This approach allows the model to more effectively represent complex scenes. Additionally, PreSem-Surf employs progressive semantic modeling, which extracts semantic information at increasing levels of detail. This strategy reduces the overall training time while improving the model’s understanding of the scene semantics. Experimental validation on seven synthetic scenes using six evaluation metrics demonstrates that PreSem-Surf achieves superior performance in metrics such as C-L1, F-score, and Intersection over Union (IoU). It also maintains competitive results in Normal Consistency (NC), Accuracy, and Completeness, highlighting its robustness and practical applicability. These results suggest that PreSem-Surf can be a valuable tool for efficient and accurate 3D surface reconstruction in applications requiring detailed scene understanding. Future work may explore extending this framework to real-world datasets and further optimizing the semantic modeling process.
👉 Pročitaj original: arXiv AI Papers
