A New Feature Coding Scheme for Video-Content Matching Tasks

Abstract

This thesis present a new feature coding scheme for video-content matching tasks. The purpose of this feature coding scheme is to compress features under a strict bitrate budget. Features contain two parts of information: the descriptors and the feature locations. We propose a variable level scalar quantizer for descriptors and a variable block size location coding scheme for feature locations. For descriptor coding, the SIFT descriptors are transformed using Karhunen-Loéve Transform (KLT). This K-L transformation matrix is trained using the descriptors extracted from the 25K-MIRFLICKR image dataset. The quantization of descriptors is applied after descriptor transformation. Our proposed descriptor quantizer allocates different bitrates to the elements in the transformed descriptor according to the sequence order. We establish the correlation between the descriptor quantizer distortion and the video matching performance, given a strict bitrate budget. Our feature location coding scheme is built upon the location histogram coding method. Instead of using uniform block size, we use different sizes of blocks to quantize different areas of a video frame. We have achieved nearly 50% reduction in the bitrate allocated for location information compared to the bitrate allocated by the coding schemes that use uniform block size. With this location coding scheme, we achieve almost the same video matching performance as that of the uniform block size coding. By combining the descriptor and location coding schemes, experimental results have shown that the overall feature coding scheme achieves excellent video matching performance.