Each clip with the extracted trajectories by gKLT tracker is preprocessed by deleting short trajectories, stationary points, and some errors.
Details of dataset can be found in dataset_info. It contains video name, video length, video size, video source, video t0 (the frame for group detection evaluation), group detection (300 group detection used in our cvpr paper), video_gt (video classes groundtruth), and scene number. (You can also choose any frame to do group detection. The frame list in video_info_t0 is just what we use in our cvpr paper.)
These data can only be used for academic research purposes.
The COPYRIGHT of the videos (with watermark) belongs to GettyImages and Pond5 respectively.
b. Dataset used in paper:
In our 14'cvpr paper, we take only 30 frames from each clip to implement and evaluate our approach.