Avenue Dataset contains 16 training and 21 testing video clips. The videos are captured in CUHK campus avenue with 30652 (15328 training, 15324 testing) frames in total.
This dataset accompanies paper "Abnormal Event Detection at 150 FPS in Matlab, Cewu Lu, Jianping Shi, Jiaya Jia, International Conference on Computer Vision, (ICCV), 2013".
Video Description
The training videos capture normal situations. Testing videos include both normal and abnormal events. Three abnormal samples are shown as follows.
| Strange action | Wrong direction | Abnormal object |
|
|
|
Our dataset contains the following challenges.
Download
Avenue Dataset (16 training videos, 21 testing videos)
Ground truth (abnormal events
marked in rectangles)
Evaluation Criteria (Spatial location)
Our evaluation not only is in frame level but also considers spatial location. We label abnormal events using rectangles. We apply the VOC Pascal style object detection criterion for our abnormal event detection.
Correct detection should satisfy: Intersection/Union > threshold.
For the frame without abnormal event, Intersection is 0. If Union = 0, we set 0/0 = 1.
We report our accuracy under thresholds {0.2,0.3,0.4,0.5,0.6,0.7,0.8} in our PAMI submission.
You can find the evaluation sample code here.