PRISMA-ORAND team: Instance Search Based on Parallel Approximate Searches

ORAND is a Chilean company focused on developing applied research in Computer Science. This paper describes the participation of the PRISMA-ORAND team for the Instance Search task (INS) at TRECVID 2012. Currently, the most common technique used to address the Instance Search problem is the codebook approach or some extension of it. In our participation we preferred a different approach which consisted in retrieving k nearest neighbors (k-NN) for each query descriptor on the full set of descriptors instead of computing a codebook. In order to efficiently solve the k-NN searches in large datasets we perform several parallel approximate searches using the open source project P-VCD and the Amazon Elastic Compute Cloud (EC2) service. We submitted four Runs to TRECVID 2012 INS task, namely: prisma-two180px, prisma-four90px, prisma-two90px, and prisma-one180px. In general, these Runs achieve satisfactory performance as their MAP are higher than the median of all the Runs in most of the topics. Additionally, we tested a fifth Run prisma-two280px by increasing both number of descriptors and search accuracy. This last Run demanded higher computational power to solve the k-NN searches but it highly improved the effectiveness compared with initial four Runs. The performance achieved by the presented approach shows some promising results. It presents an alternative to codebook approaches that can achieve satisfactory results in the Instance Search problem. The parallelization and approximation enable to solve k-NN searches in large datasets achieving high effectiveness. However, more research is needed in order to determine the potential performance that can be reached by k-NN searches compared to codebook approaches.