Submission Format. Electronic submission on Canvas is mandatory. Submit in a zip file, a single pdf file containing:
• the source code,
• the resulting disparity maps,
• the error rates from the experiments.
Also include the code and output images separately.
Problem 1. (70 points) Download the Teddy stereo pair and ground truth from the course web page and implement a winner-take all stereo algorithm using the rank-transform (see Notes 3). Compute the rank transform in 5 × 5 windows. Then, compute disparity maps on the ranktransformed images, aggregating the absolute differences of rank in 3×3 and 15×15 windows. Show the resulting disparity maps in the report. There is no need to store or show the ranktransformed images. Pixels for which any window falls out of the image boundaries can be set to black. The disparity range for these images is from 0 to 63.
Read the ground truth disparity map and divide the values by 4 and round to the nearest integer. Compute the percentage of bad pixels (error rate) by counting the fraction of pixels that differs by more than one disparity level from the ground truth (divided by 4). Differences equal to 1 are considered acceptable. Report the error rates.
1
Problem 2. (30 points) For the 3×3 aggregation window above, compute matching confidence using the PKRN measure. Using the PRKN values, generate a disparity map containing the top 50% most confident pixels. Report the error rate of the sparse disparity map and the number of pixels that have been kept. Pixels without disparity assignments should be ignored in this evaluation.
2
