A method was developed to detect generic objects using a single query image. The query image could be a typical real image, a virtual image, or even a hand-drawn sketch of the object. Without a training process, the key problem is how to describe the object class from only one query image with no pre-segmentation or other pre-processing procedures. The method introduces densely computed Scale-Invariant Feature Transform (SIFT) as the descriptor to extract “gradient distribution” features of the image. The descriptor emphasizes the edge parts and their distribution structures, which are very representative of the object class, so it is very robust and can deal with virtual images or hand-drawn sketches. Tests on car detection, face detection, and generic object detection demonstrate that the method is effective, robust, and widely applicable. The results using queries of real images compare well with other training-free methods and state-of-the-art training-based methods.

The histogram of oriented gradient has been successfully applied in many research fields with excellent performance especially in pedestrian detection. However, the method has rarely been applied to face recognition. Aimed to develop a fast and efficient new feature for face recognition, the original HOG and its variations were applied to evaluate the effects of different factors. An information theory-based criterion was also developed to evaluate the potential classification power of different features. Comparative experiments show that even with a relatively simple feature descriptor, the proposed HOG feature achieves almost the same recognition rate with much lower computational time than the widely used Gabor feature on the FRGC and CAS-PEAL databases.