Performance Analysis of 2D Feature Descriptors for Object Classification

Image descriptors play a pivotal role in image and video processing by furnishing precise descriptions of local image characteristics. These descriptors typically exhibit invariance to rotation, translation, and scaling, and in some cases, demonstrate partial invariance to illumination alterations and affine or 3D projection transformations. Unlike contemporary Machine and Deep Learning methodologies, they do not necessitate extensive datasets or substantial computational resources for application, as they operate without the need for training. Key interest points, known as keypoints, are located within images, each associated with its distinctive feature descriptor. These descriptors facilitate the comparison of features across different images, aiding in the identification of shared objects. In literature, image descriptors have primarily served in tasks such as feature matching and scene object detection. This paper aims to investigate the potential of leveraging feature descriptors for image classification tasks and to gauge their comparative performance. Specifically, we employ Support Vector Machine (SVM) and Convolutional Neural Network (CNN) classifiers to train on feature vectors extracted from images, thereby learning the local features of depicted objects. To the best of our knowledge, our study represents the pioneering effort to scrutinize the performance of top 2D descriptors for image classification. We conducted a series of experiments using widely recognized image descriptors and datasets to evaluate and compare their performances. Our objective was to ascertain the reliability of these feature descriptors for image classification tasks.