DFKI-IUPR participation in TRECVID’09 High-level Feature Extraction Task

Run No. Run ID Run Description infMAP (%) training on TV09 data (type: A) 1 IUPR-VW-TV SIFT visual words with SVMs 8.5 2 IUPR-ADAPT-TV SIFT visual words with PA1SD 5.1 combined training on YouTube and TV09 data (type: C) 3 IUPR-VW+TT-TV SIFT visual words with SVMs, fused with TubeTagger concept detection scores 8.3 4 IUPR-ADAPT-YT SIFT visual words with PA1SD, trained on YouTube, adapted to TV09 5.1 training on YouTube data (type: c) 5 IUPR-VW-YT SIFT visual words with SVMs 3.2 6 IUPR-VW+TT-YT SIFT visual words with SVMs, fused with TubeTagger concept detection scores 3.2 Similar to our TRECVID participation in 2008 [23], our main motivation in TRECVID’09 is to use web video as an alternative data source for training visual concept detectors. Web video material is publicly available at large quantities from portals like YouTube, and can form a noisy but large-scale and diverse basis for concept learning. Unfortunately, web-based concept detectors tend to be inaccurate when applied to different target domains (e.g., TRECVID data [24]). This “domain change” problem is the focus of this year’s TRECVID participation. We tackle it by introducing a highly-efficient linear discriminative approach, where a model is initially learned on a large dataset of YouTube video and then adapted to TRECVID data in a highly efficient on-line fashion. Results show that this cross-domain learning approach (infMAP 5.1%) (1) outperforms SVM detectors purely trained on YouTube (infMAP 3.2%), (2) performs as good as the linear discriminative approach trained directly on standard TRECVID’09 development data (infMAP 5.1%), but (3) is outperformed by an SVM trained on standard TRECVID’09 development data (infMAP of 8.5%).