JOANNEUM RESEARCH and Vienna University of Technology at TRECVID 2010
Publication from Digital
Gaithersburg, MD, USA Proceedings of TRECVID Workshop, 11/2010
We participated in two tasks: semantic indexing (SIN) and instance
We submitted 4 light runs, 2 with RBF kernel, 2 with a kernel combining
appropriate kernels for the different features. Two runs were trained
on the 2010 training set, two on the 2007 training set (for the 3
concepts shared between 2007 and 2010).
L A JRS-VUT1 2: RBF kernel trained on 2010 set.
L A JRS-VUT2 1: Combined kernel trained on 2010 set.
L B JRS-VUT3 4: RBF kernel trained on 2007 set.
L B JRS-VUT4 3: Combined kernel trained on 2007 set.
The combined kernel outperforms the RBF kernel on the 2010 data. For
the RBF kernel, training on 2007 data yields worse results, for the
combined kernel no clear trend can be seen.
All runs use the same features and differ by the method for fusing
and ranking results from these features.
F X NO JRS max max 4: For each shot, maximum similarity of features
of all query samples.
F X NO JRS topK 4: Top-k results for each feature (k = 1000=nFeatures).
F X NO JRS w bestR 2: Weighted linear combination of feature similarities,
weights based on best ranked other query sample.
F X NO JRS w t100 3: Weighted linear combination of feature similarities,
weights based on number of other query samples among top 100.
Features worked best for object queries, weighted fusion was better.
For persons and objects a single feature outperformed the best fused
result, for other types fused results were better than any single