PatchNeRF: learning neural radiance fields with multi-level correspondence supervision

Neural Radiance Fields (NeRF) have achieved remarkable progress in novel view synthesis and 3D reconstruction. Cross-view feature correspondences provide valuable geometric cues for improving consistency across viewpoints. In this paper, we propose PatchNeRF, a novel framework that introduces multi-level correspondence supervision into NeRF training. Specifically, we leverage point-level correspondences as geometric anchors to guide ray sampling and supervision, and further extend each corresponding point into a local patch to enforce patch-level consistency through statistical feature alignment. This design enables the network to capture both sparse geometric constraints and rich local context, leading to improved stability and structural fidelity. Experiments on complex real-world scenes demonstrate that PatchNeRF significantly enhances reconstruction quality and geometric consistency compared to existing methods.