Abstract
Previous work on disparity map fusion has mostly focused on geometric or statistical properties of disparity maps. Since failure of stereo algorithms is often consistent in many frames of a scene, it cannot be detected by such methods. Instead, we propose to use radiometric information from the original camera images together with externally supplied camera pose information to detect mismatches. As radiometric information is local information, the computations in the proposed algorithm for disparity fusion can be decoupled and parallelized to a very large degree, which allows us to easily achieve real-time performance.
Highlights
For several years, it has been possible to obtain high-quality dense distance measurements using stereo algorithms such as semi-global matching (SGM) (Hirschmüller, 2005)
The energy term consist of a fidelity term, which models the deviation from the input depth data and incorporates color information, and a regularization term, which is based on the color image based depth image denoising method proposed in (Liu et al, 2010)
We implemented a third variant (’low latency’) which differs from the basic variant in that no input views (IV) is given for the reference view (RV)
Summary
It has been possible to obtain high-quality dense distance measurements using stereo algorithms such as semi-global matching (SGM) (Hirschmüller, 2005). In stereo computation with moving or multiple cameras, it is common to fuse several disparity / depth images together to improve accuracy and to remove matching artifacts. Most methods for depth map fusion use geometric or statistical considerations to process the depth images, often without use of the information contained in the original color images. We propose a novel approach which uses color information to identify mismatched pixels, which we denote radiometry guided disparity fusion (RGDF). As color information alone provides a robust way to distinguish foreground from background, RGDF is able to eliminate mismatches even in the presence of moving objects in a scene.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
