100 likes | 239 Views
Video and stills data. Took a short video with a simple stills camera: 640x480 30fps No camera motion. Some people motion Used open-source video denoising ( Ffmpeg project) to get rid of some camera random noise. Took multiple exposures stills from (almost) the same viewpoint
E N D
Video and stills data • Took a short video with a simple stills camera: 640x480 30fps • No camera motion. • Some people motion • Used open-source video denoising (Ffmpeg project) to get rid of some camera random noise. • Took multiple exposures stills from (almost) the same viewpoint • High resolution • 3 exposures • Fused together using “Exposure Fusion”
Fused image • a better fusion or HDR composition is probably possible…
started with a single video frame. Did SIFT matching of the video frame and fused still image.
Homography fitting • Fitting an homography using RANSAC • Result on next slide – flicker back and forth
The problems • Alignment error because: • The real world geometry is not planar. • Error in SIFT-less areas. • Scene motion • How to handle people walking into the scene?
Suggested algorithm overview • For each pixel in the video frame: • If a patch around it is “clipped” (all 255 or all 1, due to extreme over/under exposure), I have no information other than the homography, take the color from the fused image according to the homography. • Otherwise, compare the patch in the video frame with patches that are up to 5 pixels from where the homography indicates. Take the best one, and take that patch’s center’s color for this pixel. If no patch is good enough – keep original color (to help keep the people in the video). • How to compare patches?
Patch comparison • Since the patches where shot at different exposures, I can’t compare color. • But Camera response curve is monotonic – i.e more light equals a brighter pixel. • Transform each patch into a “gradient direction vector”: • For each two neighboring pixels, take + if the intensity increases, - if decreases, 0 it’s the same. • A 3x3 patch is transformed into 20 trinary values. Comparison is done in this space.