Evaluation of Cost Functions for Stereo Matching
read more
Citations
Computer Vision: Algorithms and Applications
Single image dehazing
A Fast Single Image Haze Removal Algorithm Using Color Attenuation Prior
Single Image Dehazing via Multi-scale Convolutional Neural Networks
End-to-End Learning of Geometry and Context for Deep Stereo Regression
References
A taxonomy and evaluation of dense two-frame stereo correspondence algorithms
Fast approximate energy minimization via graph cuts
An experimental comparison of min-cut/max- flow algorithms for energy minimization in vision
Alignment by Maximization of Mutual Information
Fast approximate energy minimization via graph cuts
Related Papers (5)
Non-parametric local transforms for computing visual correspondence
Frequently Asked Questions (11)
Q2. What future works have the authors mentioned in the paper "Evaluation of cost functions for stereo matching" ?
Future work includes testing other matching costs that can handle radiometric differences, e. g., the census transform [ 23 ] and the approximation of MI of Zitnick et al. [ 24 ].
Q3. What are the common pixel-based matching costs?
Common pixel-based matching costs include absolute differences, squared differences, sampling-insensitive absolute differences [2], or truncated versions, both on gray and color images.
Q4. What is the effect of BT and HMI?
BT and HMI produce the best object borders, while the LoG, Rank, and especially the Mean filter cause distortions at object borders.
Q5. Why does NCC tend to blur depth discontinuities more than other matching costs?
NCC tends to blur depth discontinuities more than many other matching costs, because outliers lead to high errors within the NCC calculation.
Q6. What is the cost for correlation-based methods?
On images with simulated and real radiometric differences, the Rank transform appeared to be the best cost for correlation-based methods.
Q7. What is the way to handle the disparity images without radiometric transformations?
A qualitative evaluation of the disparity images from images without radiometric transformations indicated that the filter-based costs (LoG, Rank and Mean) tend to blur object boundaries.
Q8. What is the common method of matching images?
The simplest matching costs assume constant intensities at matching image locations, but more robust costs model (explicitly or implicitly) certain radiometric changes and/or noise.
Q9. What is the reason why the different costs perform better than BT?
For correlation the different costs perform quite similar, probably since summing over a fixed window acts like averaging, which reduces the effect of Gaussian noise.
Q10. How many different methods were evaluated for the teddy image?
Each cost was evaluated with three different stereo algorithms: a local correlation method, a semi-global matching method, and a global method using graph cuts.
Q11. What is the cost for global matching methods?
In tests with global radiometric changes or noise, hierarchical mutual information performed best for pixel-based global matching methods like SGM and GC.