MMNeRF: Multi-Modal and Multi-View Optimized Cross-Scene Neural Radiance Fields

Abstract

We present MMNeRF, a simple yet powerful learning framework for highly photo-realistic novel view synthesis by learning Multi-modal and Multi-view features to guide neural radiance fields to a generic model. Novel view synthesis has achieved great improvement with the significant success of NeRF-series methods. However, how to make the method generic across scenes has always been a challenging task. A good idea is to introduce 2D image features as prior knowledge for adaptive modeling, yet RGB features lack geometry and 3D spatial information, which causes shape-radiance ambiguity issues and lead to blurry and low-resolution results in the synthesis images. We propose a multi-modal multi-view method to make up for the existing methods. Specifically, we introduce depth features besides RGB features into the model and effectively fuse these multi-modal features by modality-based attention. Furthermore, Our framework innovatively adopts the transformer encoder to fuse multi-view features and uses the transformer decoder to adaptively incorporate the target view with global memory. Extensive experiments are carried out on both categories-specific and category-agnostic benchmarks, and the results demonstrate that our MMNeRF achieves state-of-the-art neural rendering performance.