Mining Compact Bag-of-Patterns for Low Bit Rate Mobile Visual Search
Rongrong Ji,Ling-Yu Duan,Jie Chen,Tiejun Huang,Wen Gao
DOI: https://doi.org/10.1109/tip.2014.2324291
IF: 10.6
2014-01-01
IEEE Transactions on Image Processing
Abstract:Visual patterns, i.e., high-order combinations of visual words, contributes to a discriminative abstraction of the high-dimensional bag-of-words image representation. However, the existing visual patterns are built upon the 2D photographic concurrences of visual words, which is ill-posed comparing with their real-world 3D concurrences, since the words from different objects or different depth might be incorrectly bound into an identical pattern. On the other hand, designing compact descriptors from the mined patterns is left open. To address both issues, in this paper, we propose a novel compact bag-of-patterns (CBoPs) descriptor with an application to low bit rate mobile landmark search. First, to overcome the ill-posed 2D photographic configuration, we build up a 3D point cloud from the reference images of each landmark, therefore more accurate pattern candidates can be extracted from the 3D concurrences of visual words. A novel gravity distance metric is then proposed to mine discriminative visual patterns. Second, we come up with compact image description by introducing a CBoPs descriptor. CBoP is figured out by sparse coding over the mined visual patterns, which maximally reconstructs the original bag-of-words histogram with a minimum coding length. We developed a low bit rate mobile landmark search prototype, in which CBoP descriptor is directly extracted and sent from the mobile end to reduce the query delivery latency. The CBoP performance is quantized in several large-scale benchmarks with comparisons to the state-of-the-art compact descriptors, topic features, and hashing descriptors. We have reported comparable accuracy to the million-scale bag-of-words histogram over the million scale visual words, with high descriptor compression rate (approximately 100-bits) than the state-of-the-art bag-of-words compression scheme.