Arxiv

Abstract

1. Embed diffusion model into stereo matching network
2. Adopt multi-level network for high-resolution input
3. Fuse generated depth map to reconstruct 3D human model.

Introduction

1. Sparse-view methods, which predict geometry based on appearance, cannot produce detailed human model because of lacking sufficient multiview stereo matching.

2. Continuous models are basically obtained from traditional stereo methods based on a continuous varitional formulation, which can solved by diffusion model.

3. Pipeline:

   1. Reconstruct coarse field first by using DoubleField;
   2. Render depth maps from multiple viewpoints
   3. Compute disparity flow masks
   4. Refine disparity flow with diffusion model
      • Level 1: Use CNN to extract feature maps of disparity flow masks
      • Level 2: Condition diffusion model with feature maps
   5. Fuse 3D points through interpolation.