Correlation based matching typically produces dense depth maps by calculating the disparity at each pixel within a neighborhood. This is achieved by taking a square window of certain size around the pixel of interest in the reference image and finding the homologous pixel within the window in the target image, while moving along the corresponding scanline. The goal is to find the corresponding (correlated) pixel within a certain disparity range d (d E [0,….,dmax]) that minimizes the associated error and maximizes the similarity.
In brief, the matching process involves computation of the similarity measure for each disparity value, followed by an aggregation and optimization step. Since these steps consume a lot of processing power, there are significant speed-performance advantages to be had in optimizing the matching algorithm.
The images can be matched by taking either left image as the reference (left-to-right matching, also known as direct matching) or right image as the reference (right-to-left matching, also known as reverse matching) [2].
Similarity MeasureFormula
Sum of Absolute Differences (SAD)
Zero-mean Sum of Absolute Differences (ZSAD)
Locally scaled Sum of Absolute Differences (LSAD)
Sum of Squared Differences (SSD)
Zero-mean Sum of Squared Differences (ZSSD)
Locally scaled Sum of Squared Differences (LSSD)
Normalized Cross Correlation (NCC)
Zero-mean Normalized Cross Correlation (ZNCC)
Sum of Hamming Distances (SHD)
Sum of Absolute Differences (SAD) is one of the simplest of the similarity measures which is calculated by subtracting pixels within a square neighborhood between the reference image I1 and the target image I2 followed by the aggregation of absolute differences within the square window, and optimization with the winner-take-all (WTA) strategy [1]. If the left and right images exactly match, the resultant will be zero.
In Sum of Squared Differences (SSD), the differences are squared and aggregated within a square window and later optimized by WTA strategy. This measure has a higher computational complexity compared to SAD algorithm as it involves numerous multiplication operations.
Normalized Cross Correlation is even more complex to both SAD and SSD algorithms as it involves numerous multiplication, division and square root operations.
Sum of Hamming Distances is normally employed for matching census-transformed images (can be used on images that have not been census transformed) by computing bitwise-XOR of the values in left and right images, within a square window. This step is usually followed by a bit-counting operation which results in the final Hamming distance score.
Example: Tsukuba
Left Image
Right Image
SAD Disparity Map
ZSAD Disparity Map
LSAD Disparity Map
少吃点，吃好的。
