PreSem-Surf: RGB-D Surface Reconstruction with Progressive Semantic Modeling and SG-MLP Pre-Rendering Mechanism

Yuyan Ye, Songcen Xu, Yukai Huang, Jiali Huang, Qian Weng

Abstract

We introduce PreSem-Surf, an optimization method based on the Neural Radiance Field (NeRF) framework, in a relatively short time to reconstruct high-quality surfaces from RGB-D sequences of scenes by combining RGB information, depth information, and semantic information.Specifically, we propose a novel sampling structure SG-MLP combined with PR MLP (Preconditioning Multilayer Perceptron) to pre-render voxels, which enables the model to obtain scene-related information earlier, and more effectively distinguish between noise and local details.PreSem-Surf, compared with existing models, achieves a better balance between smoothness and accuracy of reconstruction. In addition, precision-progressive semantic modeling is introduced to extract semantic information with progressive levels of accuracy, enabling the model to learn scene information while minimizing training time.The model is trained and evaluated by means of seven scenes and six metrics from synthetic datasets, allowing for comprehensive benchmarking. On average across all scenes, our model achieved the best performance in terms of C-L1, F-score, and IoU, with its performance in NC, Acc, and Comp slightly behind the best models.
BibTeX
@misc{ye2025presem,
  author = {Yuyan Ye and Songcen Xu and Yukai Huang and Jiali Huang and Qian Weng},
  title = {{PreSem-Surf: RGB-D Surface Reconstruction with Progressive Semantic Modeling and SG-MLP Pre-Rendering Mechanism}},
  year = {2025},
  eprint = {2508.13228},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.13228},
  howpublished = {arXiv},
}

Loading related content...

Figure