Project 2: Correspondence
Algorithm Description
I'lll first describe the approach I took to implementing the three functions. For match features I iterated through the features and used matlab's knnsearch to find the two nearest neighbors. Then I found the distance between them and calculated the confidence of the match using the ratio test.
For get features I used a sift-like 128 size vector. I started get features by find the patches around each interest point of width feature_width. Then I found the cells of width feature_width/4, and using the corresponding gradients (which I also broke into patches for easy readability) I constructed histograms for each cell. I tried a clumsy version of smoothing which involved adding the magnitude of a gradient to not only the nearest bin, but also adding a fraction of that gradient to the two next-closest bins. Unfortunately this didn't seem to help at all (it lost me several percentage points). After getting and normalizing the histograms I clamped all the values larger than .2 by subtracting .09 from them. (.09 was found via trial and error. Subtracting .09 worked much better -netted me several percentage points- than just setting the large values to .2). I then re-normalized and raised the feature vector to .68 which gave me good results on notre dame (final feature vector was normalized of course). I also included code for normalized image patches as well which should be currently commented out. This code does about 20 percentage points worse than using sift-like features.
For get interest points I took all of the gradients at the beginning and applied a gausian filter with standard deviation one. Then I computed the cornerness function for each pixel. If the cornerness function was too low I set it to -10 to ensure that it wasn't chosen as a top interest point. Finally I performed non-maximum suppresion by using colfilter with the max function checking if the cornerness score was a max via comparison with that result. I used the first 1000 points to get good accuracy.
Results
My final percentages were 89 on Notre Dame, 87 on Mount Rushmore, and 7 on Guadi. The Guadi image seems to be significantly different from the others. Below are pictures of my results:
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.