Compressed Histogram of Gradients: A low bit rate descriptor

Local image features have become pervasive in the areas of computer vision and image retrieval and are increasingly finding application in real-time object recognition, 3-D reconstruction, panorama stitching, robotic mapping, and video tracking. Many of these applications require large numbers of features to be transmitted or stored, which warrants low bit rate descriptors for increased scalability and efficiency. However, current feature descriptors perform well at image recognition tasks, but are very inefficient in terms of bit-rate. For example, SIFT descriptors are conventionally stored as 1024 bits and SURF feature descriptors are stored as 2048 bits. In the present invention, the novel Compressed Histogram of Gradients (CHoG) descriptors are stored as only 50 bits, yet perform on par with SIFT and SURF descriptors. This 20x reduction in size will prove vital for computer vision applications on mobile devices.