Karlsruhe Institute of Technology (KIT) at TRECVID 2009

In this paper, we present the high-level feature detection system developed by the Computer Vision for Human-Computer Interaction Lab (CVHCI) at Karlsruhe Insitute of Technology (KIT) for the TRECVID 2009 evaluation. In our previous two participations, the feature detection system relied exclusively on global features. This year, a completely new system with the focus on local low-level features has been developed. The new system supports temporal sampling as well as spatial partitioning. Local SURF descriptors are computed for grayscale images and different color spaces. The local descriptors are transformed into a more compact histogram representation using a Bag of Words approach. Color Moments and Texture Wavelets are the only two global features remaining in this new system. For each low-level feature and concept in the evaluation, a support vector machine was trained using a grid search scheme based on video-constrained cross-validation. Finally, multiple scores are fused using a simple weighted fusion approach.