VolRecon: Volume Rendering of Signed Ray Distance Functions for Generalizable Multi-View Reconstruction

• DOI: 10.48550/arxiv.2212.08067
• Publication Date: 2022-01-01
• Authors: Yufan Ren; Fangjinhua Wang; Tong Zhang; Marc Pollefeys; Sabine Süsstrunk
• Keywords: Signed distance function; Benchmark (surveying)

The success of the Neural Radiance Fields (NeRF) in novel view synthesis has inspired researchers to propose neural implicit scene reconstruction. However, most existing reconstruction methods optimize per-scene parameters and therefore lack generalizability new scenes. We introduce VolRecon, a generalizable method with Signed Ray Distance Function (SRDF). To reconstruct fine details little noise, VolRecon combines projection features aggregated from multi-view features, volume interpolated coarse global feature volume. Using ray transformer, we compute SRDF values sampled points on then render color depth. On DTU dataset, outperforms SparseNeuS by about 30% sparse achieves comparable accuracy as MVSNet full Furthermore, our approach exhibits good generalization performance large-scale ETH3D benchmark.