Denoising Diffusion for Multi-View Stereo

In this paper, we introduce a novel approach for refining depth estimation of multi-view stereo (MVS) applications by leveraging denoising diffusion models. Because recent state-of-the-art MVS methods rely on human heuristics to get a hypothesis of where the true depth lies, it is prone to produce erroneous and noisy results, particularly at the edge of the object or background. To ameliorate these issues, we present a plugin refinement process for depth estimate based on stochastic prior. Specifically, the stochastic distribution of depth maps is modeled by using the diffusion model. However, directly training depth maps on the diffusion model may introduce undesirable caveats such as holes. Thus, we use the knowledge distillation technique by using multi-view stereo output with its estimation confidence. Experimental results demonstrate the effectiveness of our proposed method compared to existing state-of-the-art MVS techniques.