TRECVID2010 Multimedia Event Detection (MED) System

Back to Project List



We participated in the Multimedia Event Detection (MED) task at NIST TREC Video Retrieval Evaluation (TRECVID) 2010. The MED task was set up to advance research and development of systems that can automatically find video clips containing any event of interest. Figure 1 gives some example frames of the three evaluated events in MED 2010. Automatic detection of such complex video events has great potential for many applications, such as web video indexing, consumer content management, and open-source intelligence analysis.

This project page introduces the techniques used in our MED 2010 system, which achieved the best performance among submissions from all participants. Figure 2 shows the evaluation results.

TRECVID is an open forum for encouraging and evaluating new research in video retrieval. It features a benchmark activity sponsored annually, since 2001, by the National Institute of Standards and Technology (NIST). More details about the evaluation procedures and outcomes can be found at the NIST TRECVID site.

Figure 1: Example frames of the 3 events evaluated in TRECVID MED 2010. Content of the same class can be very different.

Figure 2: Performance of our submitted MED runs (circled) and all 45 official submissions. The vertical axis shows the performance measured by mean Minimal Normalized Cost (MNC) of the three evaluated events. Note lower MNC means better performance.

Technical Components

In MED 2010, we aim at answering the following questions. What kind of feature is more effective for multimedia event detection? Are features from different feature modalities (e.g., audio and visual) complementary for event detection? Can we benefit from generic concept detection of background scenes, human actions, and audio concepts? Are sequence matching and event-specific object detectors critical? Our findings indicate that spatial-temporal feature is very effective for event detection, and it's also very complementary to other features such as static SIFT and audio features. As a result, our baseline run combining these three features already achieves very impressive results, with a mean MNC of 0.586. Incorporating the generic concept detectors using a graph diffusion algorithm provides marginal gains (mean MNC 0.579). Sequence matching with Earth Mover's Distance (EMD) further improves the results (mean MNC 0.565). The event-specific detector ("batter"), however, didn't prove useful from our current re-ranking tests. We conclude that it is important to combine strong complementary features from multiple modalities for multimedia event detection, and cross-frame matching is helpful in coping with temporal order variation. Leveraging contextual concept detectors and foreground activities remains a very attractive direction requiring further research.

Figure 3 shows an overview of our system, which contains four major components: 1) multiple feature modalities; 2) EMD-based temporal matching; 3) contextual diffusion; and 4) event-specific object detector. We briefly describe each of the components below.

Figure 3: Framework overview. Our system contains four major components, based on which we submitted 6 result runs.

1. Multiple Feature Modalities

We consider three visual/audio features that are expected to be useful and complementary for event detection in videos.

Static SIFT feature: We adopt two sparse keypoint detectors: Difference of Gaussian (DoG) and Hessian Affine. Since the two detectors extract keypoints of different properties, we expect that they are complementary. Using multiple keypoint detectors is also suggested by many previous works for better performance. SIFT is then adopted to describe each keypoint as a 128 dimensional vector. As processing all MED video frames will be computationally very expensive, we sample one frame from every two seconds.

Spatial-temporal interest points: While SIFT describes 2D local structures in images, spatial-temporal interest points (STIP) capture space-time volumes where the image values have significant local variations in both space and time. We use Laptev's method to compute locations and descriptors for STIPs in video. The detector is based on an extension of Harris operator to space-time. Their code does not contain scale selection; instead interest points are detected at multiple spatial and temporal scales. HOG (Histograms of Oriented Gradients; 72 dimensions) and HOF (Histograms of Optical Flow; 72 dimensions) descriptors are computed for the 3D video patches in the neighborhood of the detected STIPs. We use concatenated HOGHOF feature (144 dimensions) as the final descriptor for each STIP.

MFCC audio feature: In addition to visual features like SIFT and STIP, audio is another important cue for detecting events in videos. We expect it to be complementary to the visual features. For this, we use the popular MFCC (Mel-frequency cepstral coefficients) feature. We compute MFCC feature over every 32ms time-window with 50% (16ms) overlap.

Given a video clip, we now have three sets of features (SIFT, STIP, and MFCC). Due to the variations in length and content complexity, the sets of the same feature differ in cardinality across different video clips. We therefore adopt the popular bag-of-X representation to convert the feature sets into fixed dimensional histograms, one for each descriptor type. Details of the our feature representations can be found in our notebook paper. With the bag-of-X histogram features, we train separate one-versus-all χ2 kernel SVMs as our baseline classifiers. The average fusion of probability predictions from the three SVM classifiers forms our baseline submission run 6.

2. EMD-based Temporal Matching

While accumulating all the SIFT/STIP/MFCC features from a video clip into a single feature vector seems a reasonable choice, it neglects the temporal information within the video clip. We therefore apply the earth mover's distance (EMD) to measure video clip similarity. We only adopt SIFT feature in this experiment, and thus each video clip is represented by a sequence of 5000-d bag-of-visual-word feature vectors (cardinality equals to the number of sampled frames). As shown in Figure 4, EMD computes the optimal flows between two sets of frames/features, producing the optimal match between both sets. The EMD is then used in a generalized form of Gaussian kernel for SVM classification.

Figure 4: Toy example of EMD-based temporal matching between two sets of frames P and Q; lines indicate the presence of nonzero flows between corresponding frame pairs.

3. Contextual Diffusion

Events are mostly defined by several (moving) objects such as "person", and generally occur under particular scene settings with certain audio sounds. For example, as shown in Figure 1, "batting a run in" contains people of various actions in the baseball field scene with typically some cheering or clapping sounds. Such event-scene-object-sound dependency provides rich contextual information for understanding the events. Most previous approaches, however, handled events, scenes, objects, and audio sounds separately without considering their relationship. Our intuition is that once the contextual cues can be computed, they can be utilized to make the event detection more robust. We therefore explore such contexts in MED 2010.

To build detectors (classifiers) for a large number of contextual concepts, we first need to collect enough training samples. To this end, we defined 21 contextual concepts as listed below the correlation table in Figure 5, and designed an annotation tool to label the training videos for the 21 concepts (download annotations). Each video is divided to multiple 10-sec clips. With the labeled training data, we train SVM classifiers for detecting the concepts. For Human Action concepts, we use the bag-of-X representation of the STIP feature, while for the scene and audio concepts, we adopt similar classification framework based on the SIFT and MFCC features respectively.

To utilize these contextual detectors, we apply a contextual diffusion algorithm DASD (domain adaptive semantic diffusion) proposed in our prior work. One underlying assumption of DASD is that detectors of frequently concurrent concepts/events should produce highly correlated scores. For example, the detection result of "baseball field" and "batting a run in" should be highly consistent as they frequently co-occur. We therefore construct an undirected and weighted graph, namely semantic graph, to model the relationship between (and within) the events and the contextual concepts, where the relationship is estimated according to ground-truth labels of the events/concepts over the development dataset. Here the graph contains 24 nodes (3 events and 21 contextual concepts). A part of the graph node relationship (only the event to concept relationship) is visualized in Figure 5. The graph is then applied to refine the detection scores using a function level diffusion process, where the aim is to recover the consistency of the detection scores w.r.t. the pair-wise relationship.

Figure 5: Estimated relationship (correlation) between the 3 events and the 21 contextual concepts according to ground-truth annotations. The color-highlighted cells indicate the strong correlations discovered between events and the concepts.

4. Event-specific Object Detector

Besides the generic methods mentioned above, we are also interested in evaluating some ad-hoc ideas that are specific to individual events only. For event "batting a run in", videos usually contain certain human objects (e.g., batters) of familiar gestures and similar clothing. Assuming that videos of this event should have a high ratio of frames with batter visible, we trained a "batter" detector as an additional clue for detecting this specific event. Example detection results are given in Figure 6. The batter detector is applied as a post-processing reranking step. For every test video, the ratio of frames that have positive detection of the batter object is first computed. If the ratio is larger than an estimated threshold, the event score of the video clip will be multiplied with the ratio and the video will be moved up to the top of the ranked list. Event detection scores of other videos are not modified.

Figure 6: "batter" detection results.

For detailed result analysis of each technical component, please refer to our TRECVID 2010 paper and talk slides.


  • Yu-Gang Jiang1, Xiaohong Zeng1, Guangnan Ye1, Subh Bhattacharya2, Dan Ellis1, Mubarak Shah2, Shih-Fu Chang1
    (1Columbia University; 2University of Central Florida)



Related Projects

For problems or questions regarding this webpage please contact the Web Master.
Last updated: Dec. 2010.