TRECVID High Level Feature Extraction

About Trecvid

TRECVID encourages research in video anaylsis and retrieval by providing participants with large amounts of data and allowing groups to compare their results with others around the globe. TRECVID sponsors four different tasks, high level feature extraction, surveillance event detection, seach, and content-copy detection. The group at UCF participated in the high level feature extraction task.

About High Level Feature Extraction

In high level feature extraction, the goal is to find from over 280 hours of video specific shots that contain such high level features as "chair," "person playing soccer," "doorway," "classroom," "boat/ship," along with 15 other features specified by TRECVID. The TRECVID data set is ver challenging, containing a large amount of video with lots of variation, occlusion, and background clutter.

Our System

We implemented a basic bag of words system suing sift features and then proceeded to explore different modifications and additions to improve our results.

• SVM vs Logistic Regression (Me)
• Keyframe Extraction (Sean)
• Bootstrapping (Silvino)
• Optical Flow (Me)
• Weighting of Words (Me)
• Incorporation of GIST features (Naveed)