Efficient Object Indexing and Retrieval Using Partial Dominant Orientation Descriptor

This paper describes a method for object indexing and retrieval using a new feature descriptor called Partial Dominant Orientation Descriptor (PDOD). The extraction process of the PDOD starts by sampling the object into a set of stable and informative key locations using Difference of Gaussian (DoG), so that the retrieval can proceed successfully despite changes in object viewpoint, scale, illumination, and distortion. The proposed descriptor at feature point takes into account the position and partially computes the dominant orientations of other key locations relative to this point, thus, offering a global distinctive and discriminative characterization. The extracted object descriptors are then indexed using Vocabulary Tree, which provides a robust object retrieval system across a substantial range of rotation variance, change in textures and colors, and object deformation. The extensive experiments on KONKLAB public dataset demonstrate that our method outperforms other benchmarks such as SIFT, PCA-SIFT and SURF indexing algorithms.