TNO Instance Search at TRECVID 2010

What we learned from our runs: using a commercial facedetection package without tweaking on this (low image quality) dataset does not work. Using a small-sized (512 words) visual vocabulary computed on the query set significantly outperforms a much larger (4096 words) visual vocabulary on the whole dataset. One can build an image-retrieval system using open-source components. 2. INTRODUCTION In this notebook paper we describe our approaches to the TRECVID 2010 instance search tasks and analyze the results of our submissions. TNO has submitted three runs: two runs using a bag-of-visual-words approach and one run using a commercialoff-the-shelf (COTS) face-recognition software package. The main rationale behind all three runs was: “Can we build an instance-search system from scratch using only open source or commercial components without significant algorithmic development of our own?” The remainder of this notebook paper is as follows. The paper starts with a short section on data analysis of the video and query data set in Section 3. In Section 4 we describe in detail the processing steps of the three runs and their software implementation. In Section 5 we present some of the results of the three runs and compare them. In Section 6 we discuss some of the observations we have made on the video and query data as well as the chosen algorithms for the different runs.