Shot aggregating strategy for near-duplicate video retrieval

In this paper, we propose a new strategy for near-duplicate video retrieval that is based on shot aggregation. We investigate different methods for shot aggregation with the main objective to solve the difficult trade-off between performance, scalability and speed. The proposed short aggregation is based on two steps. The first step consists of keyframes selection. And the second one is the aggregation of the keyframes per shot. The aggregation is performed by applying Fisher vector on the descriptors computed on the selected keyframes. We demonstrate that the scalability and the speed are tackled by a sparse video analysis approach (i.e. extracting only few keyframes) combined with shot aggregation, while the performance is discussed around the choice of the aggregation strategy. The performance is evaluated on the CC_WEB_VIDEO dataset that is designed for the near-duplicate video retrieval assessment and for which some experiments have been conducted by different authors.