Heatmap from Scientific Research

Attention and corresponding hallucination of a video sequence. From top to bottom: (1) input frames, (2) attention, and (3) hallucination. Negative SSIM scores between the attention and hallucination are included at the bottom (0 means most different and -1 means most similar). Activated regions of the attentions and hallucinations match the movements of the hands across frames, showing the temporal consistency property.

DOI