Score Propagation Based on Similarity Shot Graph for Improving Visual Object Retrieval

The Visual Object Retrieval problem consists in locating the occurrences of a specific entity in an image/video dataset. In this work, we focus on discovering new occurrences of an entity by propagating the detection scores of already computed candidates to other video segments. The score propagation follows the edges of a pre-computed Similarity Shot Graph (SSG). The SSG connects video segments that are similar according to some criterion. Four methods for creating the SSG are presented: two based on computing and comparing low-level visual features, one based on comparing text transcriptions, and other based on computing and comparing high-level concepts. The score propagation is evaluated on the INS 2014 dataset. The results show that the detection performance can be slightly improved by the proposed algorithm. However, the performance is variable and depends on the properties of the SSG and the target entity. It is part of the future work to automatically decide the kind of SSG that will be used to propagate scores given a set of detection candidates.