High-resolution SVBRDF estimation based on deep inverse rendering from two-shot images

The surface of a large number of objects in the real world shows the characteristics of stationary material. These appearances usually frequently exhibit the same or similar reflectance properties in different locations, while recent methods based on deep learning can estimate the surface reflectance properties from one or multiple photographs. For stationary materials, these lightweight methods do not use the potential reflectance information to recover plausible material properties from fewer images or produce better results from the same number of images. Moreover, directly using high-resolution images to optimize models greatly increases GPU consumption. In this paper, we present a pipeline capable of reconstructing high-resolution material properties from two images taken with the flash turned on or turned off. To reduce the number of captures to two, we utilize the stationary feature to generate multiple observations of the same small region. The new observations and original observation are used to estimate the higher-quality reflectance properties of this area. We then map the estimated reflectance properties to the high-resolution appearance parameters. Optimizing the model using only small area images reduces GPU consumption. Furthermore, we use a high-resolution flash image to estimate initial high-resolution reflectance parameter maps, crop the same location area from the initial results to initialize the auto-encoder. In addition, we leverage the jump rearrangement to reduce artifacts of high-resolution results. For reconstruction, two refinements that can reintroduce details are indispensable. We demonstrate and evaluate our method on the synthetic and real data.