JOANNEUM RESEARCH and Vienna University of Technology at TRECVID 2011: Semantic Indexing and Instance Search

We participated in two tasks: semantic indexing (SIN) and instance search (INS). SIN runs We submitted one light run, using multiple kernel learning (MKL) to combine longest commmon subsequence kernels with different similarity parameters. The features are SIFT bag of features histograms and global color and texture features. The performance for some concepts is in the expected range, while the infAP score is extremely low for five of the concepts. This issue is not observed when applying the same approach to the 2011 data, and needs further investigation. INS runs We applied two approaches with quite complementary properties: One with preprocessing and indexing (based on a bagof- features (BoF) approach using Color-SIFT), and very fast query times (at most one minute), and without any preprocessing, but performing SIFT extraction and matching at query time. We submitted the following four runs: JRSVUT1: indexed Color-SIFT JRSVUT2: SIFT matching at query JRSVUT3: top results of SIFT matching at query, and indexed Color-SIFT (densely sampled) results JRSVUT4: top results of SIFT matching at query, and indexed Color-SIFT (extracted from DoG points) results The indexing method is very fast, but results are poor. The SIFT matching at query time provides good results, at or close to the best for some queries. Queries with small sample images or no distinctive visual properties yield very low performance. Fusion improves results for many queries, but removes a large number of correct hits for a few queries.