Selective Blurring

Cut video-call bandwidth by 42.4% without hurting perceived quality.

Problem

I couldn't video-call my mom in India: her 1 Mbps connection froze every call.

Impact

Cut bandwidth 42.4% and published it openly. Keeping the person sharp and blurring the rest is now standard in Zoom and Google Meet.

“Only some parts of a video frame are actually important to convey the activity being performed… our techniques can reduce the bandwidth utilization by 42.4%.”

• Selective Blurring - Video conferencing on less bandwidth

  (2012 - 2013, UIUC)

  Keep the regions of interest sharp, blur everything else, and cut the bandwidth a video call needs.

  My master's thesis, advised by Prof. Klara Nahrstedt. On a weak connection, video calls freeze, pixelate and drift out of sync. But only some parts of each frame carry the activity; the rest is background. We found those regions of interest automatically and blurred every other pixel, so each frame compresses far smaller.

  • Five ways to find the region of interest: manual selection, OpenCV object detection, Kinect depth, GPU edge detection, and DBSCAN clustering of motion
  • Parallelized on NVIDIA GPUs with CUDA so it runs in real time, frame by frame
  • Cut bandwidth use by 42.4%; user studies found the blurring did not hurt perceived quality or understanding
  • Presented to leaders from NVIDIA, Cisco and other companies
  • Why: I couldn't video-call my mom in India, whose 1 Mbps connection froze every call
  • Keeping the person sharp and blurring the background is now standard in Zoom, Google Meet and other video apps
  • Thesis: Finding regions of interest while video conferencing in bandwidth constrained environment (M.S. Computer Science, UIUC, December 2013; PDF)