Real-time Spatiotemporal Stereo Matching Using the Dual-Cross-Bilateral Grid
Skydiving video: comparison of techniques (5.9 MB)
This video shows clips from the skydiving stereo video.
Input video as red-cyan anaglyhph.
Disparity maps computed using our DCB grid on a per-frame basis.
Note the high-frequency flickering caused by per-frame processing.
Disparity maps computed using our temporal DCB grid, integrating temporal evidence over a 5-frame window.
Note the reduced flickering compared to the per-frame appraoch above.
Source video copyright 2008 Eric Deren, Dzignlight Studios.
Noisy synthetic videos: comparison of techniques (16.6 MB)
The video shows a sequence of 5 video clips, each based on one of our new synthetic stereo videos with ground truth disparities: book, street, tanks, temple and tunnel.
The individual sub-videos are (clockwise from top-left):
- Left view of input video, with additive Gaussian noise (sigma=20) to simulate thermal imaging noise of real cameras.
- Disparity maps computed using Yoon & Kweon's adaptive support weights (PAMI 2006) on a per-frame basis. Run with left-right consistency checking (see code for details).
- Disparity maps computed using our DCB grid on a per-frame basis.
- Disparity maps computed using our temporal DCB grid, integrating temporal evidence over a 5-frame window. Note that it redoces flickering and errors for most datasets.
- Disparity maps computed using our full-kernel DCB grid on a per-frame basis.
- Ground truth disparity maps for the left view.
Note that the streaks are introduced by the infilling step that follows the left-right consistency check.
The more streaks, the less consistent a technique’s disparity estimates.
Dichromatic DCB Grid: full-size images (0.5 MB)
Full-size versions of the cropped images in Figure 2 (which considers the benefit of adding a second colour axis).
Spatial-depth super-resolution: additional results (0.3 MB)
Additional results that consider the full-kernel DCB and dichromatic DCB grid for use in Yang et al.’s spatial-depth super-resolution framework.
These were not included in the paper due to space constraints.
Back to the project page.