Uploader models for video concept detection

Abstract

In video indexing, it has been noticed that a simple uploader model was able to improve the MAP of concept detection in the TRECVID Semantic Concept Indexing (SIN) task. In this paper, we explore this idea further by comparing different types of uploader models and different types of score/rank distribution. We evaluate the performance of these combinations on the best SIN 2012 runs, and explore the impact of their parameters. We observe that the improvement is generally lower for the best runs than for the weaker runs. We also observe that tuning the models for each concept independently produces a much more significant improvement.