There is no data like less data
Benjamin Elizalde, Gerald H Friedland, Howard Lei, Ajay Divakaran · 2012
Video concept detection aims to find videos that show a certain event described as a high-level concept, e.g. "wedding ceremony" or "changing a tire". This paper presents a theoretical framework and experimental evidence suggesting that video concept detection on consumer-produced videos can be performed by what we call "percepts", which is a set of observable units with Zipfian distribution. We present an unsupervised approach to extract percepts from audio tracks, which we then use to perform experiments to provide evidence for the validity of the proposed theoretical framework using the TRECVID MED 2011 dataset. The approach suggest selecting the most relevant percepts for each concept automatically, thereby actually filtering, selecting and reducing the amount of training data needed. It is show that our framework provides a highly usable foundation for doing video retrieval on consumer-produced content and is applicable for acoustic, visual, as well as multimodal content analysis.