ICPR 2018 Paper Accepted

Our work on adaptive compression for object detection algorithms has been accepted for publication at ICPR 2018.  Video compression algorithms have been designed aiming at pleasing human viewers, and are driven by video quality metrics that are designed to account for the capabilities of the human visual system. However, thanks to the advances in computer vision systems more and more videos are going to be watched by algorithms, e.g. implementing video surveillance systems or performing automatic video tagging. This paper describes an adaptive video coding approach for computer vision-based systems. We show how to control the quality of video compression so that automatic object detectors can still process the resulting video, improving their  detection performance, by preserving the elements of the scene that are more likely to contain meaningful content. Our approach is based on computation of saliency maps exploiting a fast objectness measure.

Our system pipeline. Binary saliency maps are predicted using multiple saliency maps fused with our learned model. The final binary map is shown rightmost

The computational efficiency of this approach makes it usable in a real-time video coding pipeline. Experiments show that our technique outperforms standard H.265 in speed and coding efficiency, and can be applied to different types of video domains, from surveillance to web videos.

Average Log Miss Rate on Towncenter dataset for ACF pedestrian detector for different bitrates.


This entry was posted in Uncategorized. Bookmark the permalink.

Comments are closed.