Depth-NLOS: surface detail reconstruction from depth information in non-line-of-sight imaging

Non-Line-of-Sight Imaging (NLOS Imaging) is an advanced technique designed to capture images of scenes that are not directly visible to the camera. This method leverages transient sensors to collect time-resolved signals, enabling the reconstruction of hidden scenes. Traditional NLOS imaging techniques rely solely on surface normals rather than depth information to represent implicit surfaces. This limitation results in less accurate refinement of the surface features of reconstructed objects. To address this issue, our research introduces a novel neural implicit learning approach that incorporates depth information for optimization. By integrating depth data, we achieve more precise surface reconstruction in NLOS environments. The process involves extracting depth information through the fusion of albedo plots obtained from different perspectives, which are generated by transient images and optical flow. This combined data enhances the accuracy and quality of the reconstructed images. Additionally, we introduce a depth loss component to facilitate smoother reconstruction of object surfaces while simultaneously constraining the Signed Distance Function (SDF) regression. This dual approach ensures that the reconstructed surfaces are both smooth and accurately defined. Our method has been rigorously tested on both synthetic and real datasets, and the experimental results demonstrate its superiority over existing techniques. Our approach consistently delivers high-quality reconstructions of hidden objects in various scenarios, outperforming current methods in terms of precision and detail.