Submission Format. Electronic submission on Canvas is mandatory. Submit in a zip file, a single pdf file containing:
• a description of what you did,
• the source code,
• the accuracy specified in the last step below.
Also include the source code separately. Images are not necessary, but they can be valuable debugging tools.
Harris Corner Detection and Matching Download the Teddy stereo pair and ground truth disparity map from the same link as for the second assignment. This is the only time in this course that we will IGNORE the epipolar constraint. The only reasons for doing this is that the ground truth disparity map can be used to evaluate the matched corners without any manual labeling. Implement the following steps:
1. Compute the image derivatives Ix and Iy for each pixel of image separately. Use
to compute Ix and its transpose to compute Iy. Use these first order derivaties to compute Ix2, Iy2 and Ixy at each pixel.
1
2. Apply Gaussian smoothing to the derivatives using the 5⇥ 5 filter shown at. If for some reason, youhttp:// homepages.inf.ed.ac.uk/rbf/HIPR2/gsmooth.htm are unable to do this, average the derivative values in 5⇥ 5 windows centered at each pixel. Averaging will be penalized, but a filtering step is absolutely necessary to proceed.
3. Compute the Harris operator response function. Pick as corners all pixels above a thresholdso that you select about 300-500 corners per image after the next step. 2000 is a good starting value for the threshold.
4. ApplyThis means that if a pixel does not have the maximum response in its 3then it should not be included in the output. Make sure that the order you process pixels doesnon-maximum suppression on the responses of the Harris operator in 3⇥3 neighborhood,⇥3 windows.
not affect the output of this step.
5. Compute the distance between every corner of the left image to every corner of the rightimage. IGNORE the epipolar constraint and use SAD in 3⇥3 windows for computing distances. If there are N corners in the left image and M corners in the right image there should be NM potential correspondences for which you should compute distance values.
6. Sort the distance values and select the top 5% most likely correspondences. Report thenumber of correct and incorrect correspondences included in this set. Use the ground truth disparity map and count small errors up to 1 pixel as correct. Repeat for the top 10% all the way to 100% in 5% increments.
Ignore all pixels for which any window falls out of the boundaries of the image.
2
