3D shape retrieval from a photo using intrinsic image

In recent years, 3D shape objects have spread on the Internet. Using a 2D photo as a query for 3D shape retrieval is usually much easier than preparing a 3D shape object or drawing a 2D sketch. We propose a new method for photo-based 3D shape retrieval using a so-called "Intrinsic Image." Intrinsic Image enables us to separate a given 2D photo into "Reflectance" and "Shading" images. We have observed that during the separation, texture information is primarily captured by "Reflectance," while shape information is left within "Shading." After the separation, we employ Histogram of Oriented Gradients (HOG) to extract the feature vector from "Shading" images, and apply principal component analysis (PCA) to obtain robustness against rotation, which has been the biggest problem of HOG. We conducted experiments with a commonly available 3D shape benchmark, compared our proposed method with the previous methods, and demonstrated that our method outperformed them in terms of 1st-Tier, 2nd-Tier, and P@1.