In this paper we describe our TRECVID 2011 video retrieval experiments. The MediaMill team participated in two tasks: semantic indexing and multimedia event detection. The starting point for the MediaMill detection approach is our top-performing bag-of-words system of TRECVID 2010, which uses multiple color SIFT descriptors, sparse codebooks with spatial pyramids, and kernel-based machine learning. All supported by GPU-optimized algorithms, approximated histogram intersection kernels, and multi-frame video processing. This year our experiments focus on 1) the soft assignment of descriptors with the use of difference coding, 2) the exploration of bag-of-words for event detection, and 3) the selection of informative concepts out of 1,346 concept detectors as a representation for event detection. The 2011 edition of the TRECVID benchmark has again been a fruitful participation for the MediaMill team, resulting in the runner-up ranking for concept detection in the semantic indexing task.

In this paper we describe our TRECVID 2009 video retrieval experiments. The MediaMill team participated in three tasks: concept detection, automatic search, and interactive search. The starting point for the MediaMill concept detection approach is our top-performing bag-of-words system of last year, which uses multiple color descriptors, codebooks with soft-assignment, and kernel-based supervised learning. We improve upon this baseline system by exploring two novel research directions. Firstly, we study a multi-modal extension by including 20 audio concepts and fusion using two novel multi-kernel supervised learning methods. Secondly, with the help of recently proposed algorithmic refinements of bag-of-word representations, a GPU implementation, and compute clusters, we scale-up the amount of visual information analyzed by an order of magnitude, to a total of 1,000,000 i-frames. Our experiments evaluate the merit of these new components, ultimately leading to 64 robust concept detectors for video retrieval. For retrieval, a robust but limited set of concept detectors justifies the need to rely on as many auxiliary information channels as possible. For automatic search we therefore explore how we can learn to rank various information channels simultaneously to maximize video search results for a given topic. To further improve the video retrieval results, our interactive search experiments investigate the roles of visualizing preview results for a certain browse-dimension and relevance feedback mechanisms that learn to solve complex search topics by analysis from user browsing behavior. The 2009 edition of the TRECVID benchmark has again been a fruitful participation for the MediaMill team, resulting in the top ranking for both concept detection and interactive search. Again a lot has been learned during this year's TRECVID campaign; we highlight the most important lessons at the end of this paper.

In this paper we describe our TRECVID 2008 video retrieval experiments. The MediaMill team participated in three tasks: concept detection, automatic search, and interactive search. Rather than continuing to increase the number of concept detectors available for retrieval, our TRECVID 2008 experiments focus on increasing the robustness of a small set of detectors using a bag-of-words approach. To that end, our concept detection experiments emphasize in particular the role of visual sampling, the value of color invariant features, the influence of codebook construction, and the effectiveness of kernel-based learning parameters. For retrieval, a robust but limited set of concept detectors necessitates the need to rely on as many auxiliary information channels as possible. Therefore, our automatic search experiments focus on predicting which information channel to trust given a certain topic, leading to a novel framework for predictive video retrieval. To improve the video retrieval results further, our interactive search experiments investigate the roles of visualizing preview results for a certain browse-dimension and active learning mechanisms that learn to solve complex search topics by analysis from user browsing behavior. The 2008 edition of the TRECVID benchmark has been the most successful MediaMill participation to date, resulting in the top ranking for both concept detection and interactive search, and a runner-up ranking for automatic retrieval. Again a lot has been learned during this year's TRECVID campaign; we highlight the most important lessons at the end of this paper.

In this paper we describe our TRECVID 2007 experiments. The MediaMill team participated in two tasks: concept detection and search. For concept detection we extract region-based image features, on grid, keypoint, and segmentation level, which we combine with various supervised learners. In addition, we explore the utility of temporal image features. A late fusion approach of all region-based analysis methods using geometric mean was our most successful run. What is more, using MediaMill Challenge and LSCOM annotations, our visual-only approach generalizes to a set of 572 concept detectors. To handle such a large thesaurus in retrieval, an engine is developed which automatically selects a set of relevant concept detectors based on text matching, ontology querying, and visual concept likelihood. The suggestion engine is evaluated as part of the automatic search task and forms the entry point for our interactive search experiments. For this task we experiment with two browsers for interactive exploration: the well-known CrossBrowser and the novel ForkBrowser. It was found that, while retrieval performance varies substantially per topic, the ForkBrowser is able to produce the same overall results as the CrossBrowser. However, the ForkBrowser obtains top-performance for most topics with less user interaction. Indicating the potential of this browser for interactive search. Similar to previous years our best interactive search runs yield high overall performance, ranking 3rd and 4th.

In this paper we describe our TRECVID 2006 experiments. The MediaMill team participated in two tasks: concept detection and search. For concept detection we use the MediaMill Challenge as experimental platform. The MediaMill Challenge divides the generic video indexing problem into a visual-only, textual-only, early fusion, late fusion, and combined analysis experiment. We provide a baseline implementation for each experiment together with baseline results, which we made available for the TRECVID community. The Challenge package was downloaded more than 80 times and we anticipate that it has been used by several teams for their 2006 submission. Our Challenge experiments focus specifically on visual-only analysis of video (run id: B_MM). We extract image features, on global, regional, and keypoint level, which we combine with various supervised learners. A late fusion approach of visual-only analysis methods using geometric mean was our most successful run. With this run we conquer the Challenge baseline by more than 50%. Our concept detection experiments have resulted in the best score for three concepts: i.e. desert, flag us, and charts. What is more, using LSCOM annotations, our visual-only approach generalizes well to a set of 491 concept detectors. To handle such a large thesaurus in retrieval, an engine is developed which automatically selects a set of relevant concept detectors based on text matching and ontology querying. The suggestion engine is evaluated as part of the automatic search task (run id: A-MM) and forms the entry point for our interactive search experiments (run id: A-MM). Here we experiment with query by object matching and two browsers for interactive exploration: the CrossBrowser and the novel NovaBrowser. It was found that the NovaBrowser is able to produce the same results as the CrossBrowser, but with less user interaction. Similar to previous years our best interactive search runs yield top performance, ranking 2nd and 6th overall. Again a lot has been learned during this year's TRECVID campaign, we highlight the most important lessons at the end of this paper.

In this paper we describe our TRECVID 2005 experiments. The UvA-MediaMill team participated in four tasks. For the detection of camera work (runid: A_CAM) we investigate the benefit of using a tessellation of detectors in combination with supervised learning over a standard approach using global image information. Experiments indicate that average precision results increase drastically, especially for pan (+51%) and tilt (+28%). For concept detection we propose a generic approach using our semantic pathfinder. Most important novelty compared to last years system is the improved visual analysis using proto-concepts based on Wiccest features. In addition, the path selection mechanism was extended. Based on the semantic pathfinder architecture we are currently able to detect an unprecedented lexicon of 101 semantic concepts in a generic fashion. We performed a large set of experiments (runid: B_vA). The results show that an optimal strategy for generic multimedia analysis is one that learns from the training set on a per-concept basis which tactic to follow. Experiments also indicate that our visual analysis approach is highly promising. The lexicon of 101 semantic concepts forms the basis for our search experiments (runid: B_2_A-MM). We participated in automatic, manual (using only visual information), and interactive search. The lexicon-driven retrieval paradigm aids substantially in all search tasks. When coupled with interaction, exploiting several novel browsing schemes of our semantic video search engine, results are excellent. We obtain a top-3 result for 19 out of 24 search topics. In addition, we obtain the highest mean average precision of all search participants. We exploited the technology developed for the above tasks to explore the BBC rushes. Most intriguing result is that from the lexicon of 101 visual-only models trained for news data 25 concepts perform reasonably well on BBC data also.

This year the UvA-MediaMill team participated in the Feature Extraction and Search Task. We developed a generic approach for semantic concept classification using the semantic value chain. The semantic value chain extracts concepts from video documents based on three consecutive analysis links, named the content link, the style link, and the context link. Various experiments within the analysis links were performed, showing amongst others the merit of processing beyond key frames, the value of style elements, and the importance of learning semantic context. For all experiments a lexicon of 32 concepts was exploited, 10 of which are part of the Feature Extraction Task. Top three system-based ranking in 8 out of the 10 benchmark concepts indicates that our approach is very promising. Apart from this, the lexicon of 32 concepts proved very useful in an interactive search scenario with our semantic video search engine, where we obtained the highest mean average precision of all participants.