1. Introduction
Current non-line-of-sight (NLOS) imaging techniques typically adopt pulse lasers and time-resolved detectors to image hidden objects behind obstacles or around corners, as illustrated in Fig. 1(a). To reconstruct the hidden objects from the detector’s measurements, known as NLOS transients, most physics-based NLOS imaging methods, such as Light-cone transform (LCT) [25], assume that the NLOS scene only involves three-bounce reflections and has no self-occlusion, which simplifies the problem into a linear one in the Fourier domain. However, these assumptions may not hold for more complex objects with large depth variations, which are common in practical NLOS tasks, such as tilted vehicles in autonomous driving. In these scenarios, the invalid assumptions of most physics-based methods make it challenging to recover the high-frequency details of the hidden objects. Moreover, studies have found that objects with large depth variations in NLOS problems typically have complex geometries and normal distributions, which may result in a loss of high-frequency information in the Fourier domain [17] due to the limited NLOS aperture [18], [22], rendering the NLOS imaging problem highly ill-posed.