Orand Team: Instance search and multimedia event detection using k-NN searches

ORAND S.A. is a Chilean company focused on developing applied research in Computer Science. This report describes the participation of the ORAND team at Instance Search task (INS) and Multimedia Event Detection task (MED) in TRECVID 2013. The INS participation considered four submissions, namely: orand-1sift, orand-2sift, orand-graph, and orand-interactive. The first two submissions follow the approximate k-NN search approach we presented at TRECVID 2012, the last two submissions use a static similarity graph between shots to propagate scores. In general, our submissions achieve satisfactory performance: their MAP are higher than the median in every topic, and they achieved the highest MAP in two topics. The MED participation considered one submission to pre-specified events and one submission to ad-hoc event detection. The first submission follows a naive BOW approach and achieved low performance. The second submission follows the approximate k-NN search approach and achieved higher performance. This is our first participation at MED, hence we still need more work in order to achieve competitive performance in this task.