Abstract
Neural implicit surface reconstruction has achieved impressive results using differentiable volume rendering, yet state-of-the-art methods like NeuS [20] often fail on complex, textureless objects and intricate thin structures (e.g., insects) where photometric consistency is ambiguous. To address this, we present Sem-NeuS, a framework that leverages robust semantic priors from Vision Foundation Models (DINOv3 [18]) to guide geometric optimization. Our approach combines a hierarchical coarse-to-fine Signed Distance Function (SDF) with a novel Semantic-Guided Geometry Distillation strategy. Rather than using semantic features as input, we treat them as a ground-truth supervision signal: a lightweight head is trained to predict multi-view consistent semantic embeddings solely from the underlying geometry. This effectively resolves ambiguities in low-texture regions without compromising physical validity. Experiments on the OmniObject3D benchmark and on a custom entomological dataset demonstrate that Sem-NeuS significantly outperforms state-of-the-art methods in reconstruction fidelity and normal consistency for intricate structures. The dataset and code are available at: https://github.com/amrani400/Sem-NeuS.