Hey buddy. You have a really nice explanation style and it comes from the fact that you have a very in depth understanding of the subject matter. This was my first time learning SIFT and I understood it completely. Good job. If you continue to post technical content like this, ill definitely subscribe. Hope your interest in computer vision stays the same!
I don't understand why we do scale space extrema detection and not just space extrema detection. In a previous slide, you show a 1D example with convolution with the laplacian of Gaussian, but there are not several values of sigma in that slide...
If you see the slide titled (Coming to the point) you can see that the concept works only when the size of the blob is similar to that of the sigma value of the laplacian. Hence its important to serach within a range of different sigma values
Loved the way you explained it. Thanks a lot. I have one question. In scale-space extrema detection, do we need to always compare the middle pixel of the second(intermediate image)? I don't understand that part.
Thank you for your feedback. 😇 To answer your question, it's not necessarily the second image it can be any image from the second image to second last image. Basically the pixels should have 26 neighbours in total. 9 above 9 below and 8 in the same plane.
Wow quite lucidly explained! At 28:05 how does the transpose of the inverse multiplied with the original matrix become identity matrix? Inverse multiplied with the matrix itself gives an identity right? Anyway thanks for the video!
Gaussian simply blurs the image (or we can say cancels the white noise). Besides, derivate of gaussian determines the changes in pixel values/edges. Therefore, for detecting edges it's necessary to use the derivative of gaussian.
Thanks for the nice video. I have a doubt in the step 'scale-space extrema detection'. For an octave: (considering 5 different scales of images created using Gaussian blur), we would be having 4 resulting DoG images from the previous step. So it's understandable to compare pixels from the 2nd DoG image with its neighbors from the 1st and the 3rd DoG images. Similarly, we could compare pixels from the 3rd DoG image with the 2nd and the 4th images. But how about the pixels in 1st and the 4th (topmost in that octave)? With whom should those be compared? Or we just consider only those in the middle (2nd and 3rd from the four DoG images from the previous step!!)
We only consider those that are in middle . For more better understanding you can see the description i have put a link of visual interaction and explanation of algorithm please check that.
