Abstract:  The appearance of the real world primarily depends on such intrinsic properties of images as the geometry of objects in the scene, the surface material, and the direction and intensity of illumination. Predicting these intrinsic properties from two-dimensional images is a classical problem in computer vision and graphics, and is of great importance in three-dimensional image reconstruction and augmented reality applications. However, the prediction of intrinsic properties of two-dimensional images is a high-dimensional and ill-posed inverse problem, and fails to yield the desired results with traditional algorithms. In recent years, with the application of deep learning to various aspects of two-dimensional image processing, a large number of research results have predicted the intrinsic properties of images through deep learning. The algorithm framework was proposed for deep learning-based image intrinsic property prediction. Then, the progress of domestic and international research was analyzed in three areas: intrinsic image prediction based on acquiring scene reflectance and shading map, intrinsic properties prediction based on acquiring material BRDF parameters, and intrinsic properties prediction based on acquiring illumination-related information. Finally, the advantages and disadvantages of each method were summarized, and the research trends and focuses for image intrinsic property prediction were identified. 

Key words: computer vision, computer graphics, intrinsic properties prediction, intrinsic image prediction, BRDF prediction, illumination prediction, deep learning