MVD-NeRF: Resolving Shape-Radiance Ambiguity via Mitigating View Dependency

We propose MVD-NeRF, a method to recover high-fidelity mesh from neural radiance fields(NeRFs). The phenomenon of shape radiance ambiguity, where the radiance of a point changes significantly when viewed from different angles, leads to incorrect geometry. We propose directed view appearance encoding to achieve the desired viewpoint invariance by optimizing the appearance embedding for each input image, enabling the network to learn the shared appearance representation of the entire image set. Moreover, we utilize skip connections and layer scale module in the network architecture to capture complex and multifaceted scene information. The introduction of layer scale module allows the shallow information of the network to be transmitted to the deep layer more accurately, maintaining the consistency of features. Extensive experiments demonstrate that, despite the inevitability of shape-radiance ambiguity, the proposed method can effectively minimize its impact on geometry which essentially mitigates the impact of view-dependent variations. The extracted geometry and textures can be deployed in any traditional graphics engine for downstream tasks.