Latent Concept Extraction for Zero-Shot Video Retrieval
Kazuya Ueki · 2018
For a two-year period, during 2016 and 2017, we tackled an ad-hoc video search (AVS) task, which is one of the research tasks in the TRECVID benchmark. The goal of this AVS task is to conduct a detailed video retrieval from a large-scale video database using a query phrase. We achieved the top accuracy for the second consecutive year by combining a large number of pre-trained concept classifiers. However, the best performance of the AVS system remains very low at around 20% and appropriate video sequences cannot be retrieved at all for certain specific query phrases. In this paper, we investigate whether such an extremely poor performance can be improved in the following two ways: (1) unveiling latent concepts from existing concept classifiers, and (2) combining multiple concept classifiers obtained using visual features.