Where do we find the Trucco and Verri book? In homework 4, it seems to suggest that we can find it at http://www.cs.otago.ac.nz/research/vision/...ow/opticalflow.html , but I don't see it there. Is an electronic version available somewhere?

At /m83, if your results look ok in general (e.g. there aren't lots of corner points right on top of each other) it should be ok.

You could use a similar algorithm, but instead loop densely through each pixel, look at a neighborhood centered around each pixel, and set the center pixel as a corner only if it is greater than all pixels in that neighborhood.

At /m84, there are some links on the course website in the reading section next to the lecture slides.

In homework 4, it says to include in the report "Any modifications to the window size and/or thresholds you used." Does this mean that we are expected to find window size/threshold values that produce better results than the specified values? Or, does it mean that we include this in the report only if the specified values don't produce good results?

I ask because I noticed that some of the specified window sizes and thresholds don't seem to produce very good results (i.e. large window sizes for corner detection).

do we need to do something special when we have color images? it seems like the equations for b in the lecture slide require a single dimensional image (grayscale), is this the case?

At /m87, you're right that corner detection will usually look more accurate for the smaller size windows, whereas sometimes the optical flow direction can be more accurate for the larger windows, and also the values of tau are probably not ideal for all images. It would be nice for people to try a couple other values for each image to try to get nice looking results, but it's not going to be that important toward grading.

At /m88, you can just convert to grayscale

Edited by author 12-03-2009 01:27 AM
Should we be normalizing our images to have intensity values in the range [0,1]?

I = double(imread('image.png'))/255;

I have to change my tao value to get the same results for the normalized and non-normalized case.

At /m91, yes it's a good idea to normalize to be in [0,1]

For homework 4, the provided derivations of the determinant and matrix inverse are merely suggestions, correct? We can just use the MATLAB 'det' function and do A\b for solving the system of equations instead, right?

At /m93, yes it's fine to use matlab's internal functions

Edited by author 12-04-2009 02:32 PM
Just a clarification, in part C, it says "Show 4 quiver plots on the same images from Part B, displaying a flow vector at each corner point."

So, for this section, we are ONLY including 4 quiver plots, and they are with whatever tao and window size we decide gives us the best results, correct? This is not with the tao and window sizes from parts A or B.

Additionally, should we use the same window size for both the corner detection and optical flow? Because it seems you'd get better results if you used a small window for corner detection to detect more corners, and then a larger window for optical flow.

At /m95, it would be good to show 4 quiver plots for each value of tau, but it's also ok to pick which one works best for you.

In the sample final, it says to show NCC and NSSD are equivalent. I haven't been able to do that, by expanding the NSSD expression, I get NSSD = -2*NCC + 2 , pretty close, but not algebraically equivalent. Is it actually possible to make them algebraically equivalent?

At /m98, I think this is just a badly worded question. I don't think they are actually algebraically equivalent, but are "equivalent" in the sense that they will give the same answer if used by a matching algorithm, since if they differ by only a constant factor the closest match will be the same. I don't think they can really be the same, since NSSD is a distance function (you want to find the match with the minimum NSSD), while NCC is a similarity function (you want to find the match with the maximum NCC)