Scene 1

1. Visualization of the predicted features from MPED-RNN.

We visualize the input features and generated features in three domains of one individual skeleton trajectory in the test video.
The deviation of generated features from input reflects its level of anomaly.
Corresponding to Figure 5 on the manuscript.

Left: Combined features. Center: Global component. Right: Local component
Red: predicted, Black: Input

The videos may not play properly in non-chrome browsers. In that case, please refer to them in "videos" folder.

2. Comparison of anomaly score maps from MPED-RNN and from other state-of-the-art methods

MPED-RNN gives scores on skeleton instead of pixels. The connection between joints are added for visibility.
Corresponding to Figure 4 on the manuscript.
Top left: Input, Top right: MPED-RNN, Bottom-left: Conv-AE [14], Bottom right: Liu et al. [20]

The videos may not play properly in non-chrome browsers. In that case, please refer to them in "videos" folder.