Model-Guided Iterative Diffusion Sampling for NLOS Reconstruction
Xiongfei Su, Yu Hong, Juntian Ye, Feihu Xu, Xin Yuan
Abstract
We consider the reconstruction problem of non-line-of-sight (NLOS) imaging. The proposed diffusion-based method starts with back-projection from NLOS measurement and iteratively estimates the noise using a trained light-weight U-Net and model-guided iterative gradient descent. Unlike previous large networks, our diffusion-based method relies on a series of iterative refinement steps, each of which is trained with a regression loss. This benefits our iterative approach to capture richer distributions. Furthermore, rather than estimating the posterior mean value, our method generates samples from the target posterior and allows only one trained model being adaptive to different settings, such as diverse sampling time resolution, various spatial resolution and multiple channels of color scenes. Simulation and real-data experiments verify that our proposed method achieves better reconstruction results both in quality and quantity than existing methods.