An empirical evaluation of translational and rotational invariance of descriptors and the classification of flower dataset

Object recognition and identification is used in the development of automatic systems in various domains. Latest research indicates that the performance of such systems depend on the efficiency in feature extraction; robust feature description and optimized classification or matching. This paper presents an empirical evaluation of efficiency and robustness of various gradient and binary descriptors with respect to translation, rotation and scaling etc. The performance of each descriptor is evaluated against the parameters such as size of feature set in terms of number of keypoints, matching accuracy and execution time. The detailed experiments were conducted on 17 category Oxford flower dataset to evaluate the robustness of descriptors against various rotations, scaling and noise using precision and recall values. Experimental results shows that the PCA-SIFT and SURF gives less matching rate but faster as compared to SIFT due to reduction in dimension in PCA-SIFT and use of integral images in SURF. ORB gives the best classification and outperforms the other descriptors with less memory requirement and is compact in size.