Near-duplicate video retrieval in a real-time manner is important to offer efficient storage services, and becomes more challenging due to dealing with the rapid growth of multimedia videos. Existing work fails to efficiently address this important problem due to overlooking the storage property of massive videos. In order to bridge the gap between storage system organization and application-aware videos, we propose a cost-effective real-time video retrieval scheme, called FastVR, which supports fast near-duplicate video retrieval. FastVR has the salient features of space- and time-efficiency in large-scale storage systems. The idea behind FastVR is to leverage space-efficient indexing structure and compact feature representation to facilitate keyframe based matching. Moreover, in the compact feature representation, FastVR transforms the frames into feature vectors in the Hamming space. The indexing structure in FastVR uses Locality Sensitive Hashing(LSH) to support fast similar neighboring search by grouping similar videos together. The conventional LSH unfortunately causes space inefficiency that is well addressed by a cuckoo hashing scheme. FastVR uses a semi-random choice to improve the performance in the random selection of the cuckoo hashing scheme. We implemented FastVR and examined the performance using a real-world dataset. The experimental results demonstrate the efficiency and significant performance improvements.
展开▼