As a French student, i understand most of the video and find it way clearer than the orignal paper from Lowe. I think it is due to to quality of the presentation, the fluency of the teacher. Moreover you can feel that the teacher knows what he is talking about! :) Great video!
Thanks to Dr Shah and the uploader. I just did this in class at my university, but it wasn't half as clear as this one. Very helpful. Amazing that an 8 year old recorded lecture is more relevant than a current live one
The part where he explained how the Laplacian of Gaussian works as a specific size of blob detector to achieve scale invariance at 18:19 was really helpful for me. My CV professor just skipped straight to difference of Gaussians and I didn't get why we used them or the benefit of it until now.
and, in the figure, two Probability Density Functions (pdf) are shown. the first bell-shape curve is the PDF for all correct cases. You can see that it's a normal distribution, or gaussian distribution. in most cases, the ratio (horizontal axis) has a value less then 0.8. The 2nd PDF is for the wrong cases, meaning we found a close but wrong match, it's a parabola, and most cases have a ratio larger than .08. so 0.8 is a good value to use.
this is for the purpose of robustness. for a descriptor in image 1, we may find more than one pretty close matches in image 2, the closeness of these matches are measured by their Euclidean distance from the descriptor in image 1. The smaller the distance the better. The ratio is the ratio between the best match and 2nd best match.
Time is NECESSARILY possible/potential AND actual IN BALANCE, AS E=MC2 IS F=ma; AS ELECTROMAGNETISM/energy is gravity ON BALANCE. Great !!!! By Frank DiMeglio
A way to downscale an image is by taking + skipping one row/col. Assuming your img is 4x4. You take Rows/Cols 1,3 while discarding Rows/Cols 2,4. You new img is 2x2 which is a downscaled version of the original. There are other methods thou. i.e. Taking the average of an 2x2 window into a pixel.
Sry maybe its a stupid question. I wonder if all the computation in the videos, where they show how they track an object like a card, is done in "real time"? I mean its computationally expensive to compute the coeffs for e.g. an affine transformation, right?
i dont understand one thing in 0:26.55. after half-sampling the image with the k^2*sigma scale, are we gonna apply gaussian with k^2*sigma scale on the image that we obtained from sampling? i would be glad if somebody explain that.
I am just wondering why Harris keypoint detector +SIFT desciptor is popular approach ? SIFT keypoints are scale invariant wheras Harris Keypoints are not ...
0.8 means first and 2nd best matches are too much closer. actual match can be 2nd best but due to some noise we are getting it as 2nd best instead of 1st or in other words 1st best can be wrong so we are taking chance. Graph is experimental results that from 0.1 to 0.8 first best match is best there are some very small wrong matches. but after 0.8 the first best is not correct. It is taken according to experimental results not according to some specific theory.
I think that when he tried to explain the Key Point matching at 57.50 he didn't see the words "minimum Euclidean Distance". That would have helped him a lot. It happens sometimes.
Dr. Shah is more concerned about citations than SIFT. I wish he described SIFT as detailed as the importance of citations (in his universe). This way there are still unexplained things.
This professor has a talent for explaining things clearly and concisely.
or maybe good slides
Or maybe both
true
I think this guy is great. This is the first time I have bothered to write something about anything on the internet apart from facebook.
One of the best lectures I have seen. Very clear explanation of all the technical steps.
As a French student, i understand most of the video and find it way clearer than the orignal paper from Lowe. I think it is due to to quality of the presentation, the fluency of the teacher. Moreover you can feel that the teacher knows what he is talking about! :) Great video!
Thanks to Dr Shah and the uploader. I just did this in class at my university, but it wasn't half as clear as this one. Very helpful. Amazing that an 8 year old recorded lecture is more relevant than a current live one
The part where he explained how the Laplacian of Gaussian works as a specific size of blob detector to achieve scale invariance at 18:19 was really helpful for me. My CV professor just skipped straight to difference of Gaussians and I didn't get why we used them or the benefit of it until now.
A polished lecture given by a nice guy. Dr. Mubarak describe SIFT in a straightforward way.
36:00 Difference between edges and interest points in terms of Laplacian of Gaussian.
BEST VIDEO ON SIFT! Explains the algorithm really well. Thank you so much.
This is the definition of greatness
The best SIFT explanation I ever found. Thanks
Even though i am new to CV he clearly made me to understand about SIFT.. Thanks! professor.. :)
Where are we using DOG's calculated on downsampled image?
after playing it almost 5 times over a month every thing is clear now.
This lecture really helped me acquire a better understanding of the SIFT algorithm. Thank you very much.
Very clear explanation! I was very interested in this topic because of the way it was delivered.
Thank you so much for these videos, very detailed and helpful :) please don't stop posting these lectures.
Does the scale refer to Sigma of Gaussian within an octave or downsampled image size?
28:55 , isnt sigma supposed to be = 1.6 in the start ( and not 0.707 )
Amazing Lecture! A comprehensive explanation
[01:08:34] «You have to write good papers which can be cited» - Dr. Mubarak Shah
That is a great demonstration on the SIFT algorithm. Thanks much!
Thank you for your contribution, it's much easier for me than reading the paper myself.
Clear and concise explanation. Smart way
It's really helpful for getting the gist of SIFT. Thank you so much!
Thanks! That's very clear explanation of SIFT. Much better than my professor..
and, in the figure, two Probability Density Functions (pdf) are shown. the first bell-shape curve is the PDF for all correct cases. You can see that it's a normal distribution, or gaussian distribution. in most cases, the ratio (horizontal axis) has a value less then 0.8. The 2nd PDF is for the wrong cases, meaning we found a close but wrong match, it's a parabola, and most cases have a ratio larger than .08. so 0.8 is a good value to use.
This lecture is so good. I loved the way of explaining it by Dr. Shah
Thanks a lot Dr. Mubarak Shah.
Extremely good and clear explanation, thank you for this video
53:50
SIFT in a nutshell ..... why SIFT is 128 dimensions and how it's extracted from actual image.
- Thanks for uploading the video :)
Thanks! I like how at 54:14 he says "and that's it. you can describe this in one slide".
amazing, very clear explanation of each step involved. good job sir
The professor looks like Mohammad Reza Pahlavi :))
Excellent lecture btw
I have to say, this point and explanation is much better than mine.
Very informative. Best explanation about SIFT
this is for the purpose of robustness. for a descriptor in image 1, we may find more than one pretty close matches in image 2, the closeness of these matches are measured by their Euclidean distance from the descriptor in image 1. The smaller the distance the better. The ratio is the ratio between the best match and 2nd best match.
Excellent video! Thanks! Keep them coming please.
Thank you sir for giving this lecture. It helps me a lot.
Time is NECESSARILY possible/potential AND actual IN BALANCE, AS E=MC2 IS F=ma; AS ELECTROMAGNETISM/energy is gravity ON BALANCE.
Great !!!!
By Frank DiMeglio
By the way. Very good lecture, thanks a lot for publishing this. God bless you all.
At 26.30 time, What does it mean by every other rows and every other column in down sampling process?
A way to downscale an image is by taking + skipping one row/col.
Assuming your img is 4x4. You take Rows/Cols 1,3 while discarding Rows/Cols 2,4. You new img is 2x2 which is a downscaled version of the original.
There are other methods thou. i.e. Taking the average of an 2x2 window into a pixel.
+Hans Hardmeier thank you lot
I've been looking for an explanation of how sigma values are computed to lead to those results for a while now. Thank you.
Demo Software: SIFT Keypoint Detector
David Lowe
www.cs.ubc.ca/~lowe/keypoints/
Can we get the presentation slides, the link provided is showing an error
Sry maybe its a stupid question. I wonder if all the computation in the videos, where they show how they track an object like a card, is done in "real time"?
I mean its computationally expensive to compute the coeffs for e.g. an affine transformation, right?
Good explanation professor.
Thank you Sir for explaining it so clearly and in great detail.
lots of thanks, Great and simple explaination
ALLL professors explain SIFT as a literature review, no one can explain it practically, we need David Lowe himself to explain his theory!
Very good lecture, helped me a lot. Thank you!
very nice lecture!
by using sift algorithm can we identify color of image?
Sir, very nice. Great lecture.
Thank you sir, for your nice explanation and information.
great series!
very grateful :)
Life-saver, thank you so much!
i dont understand one thing in 0:26.55. after half-sampling the image with the k^2*sigma scale, are we gonna apply gaussian with k^2*sigma scale on the image that we obtained from sampling? i would be glad if somebody explain that.
very good lecture,thumbs up!
Hairless points? What is he saying?
watch the previous lecture. He covered corner detection (Harris point) there
yeh those points have to be well groomed you know
I am just wondering why Harris keypoint detector +SIFT desciptor is popular approach ? SIFT keypoints are scale invariant wheras Harris Keypoints are not ...
+Sai Manoj Prakhya because they combine the qualities of descriptor and detector
Thanks a lot! Very good and detailed explanation!
very well explained
0.8 means first and 2nd best matches are too much closer. actual match can be 2nd best but due to some noise we are getting it as 2nd best instead of 1st or in other words 1st best can be wrong so we are taking chance. Graph is experimental results that from 0.1 to 0.8 first best match is best there are some very small wrong matches. but after 0.8 the first best is not correct. It is taken according to experimental results not according to some specific theory.
Can somebody explain the figure at 01:04:46 ? What is the ratio of distance from.... and what the figure says? why we choose 0.8 ?
is there any lecture for SURF ? ... similar to SIFT
wonderful lecture ...Thanks a lot
can anyone explain zero crossings vs scale space graph.
I think that when he tried to explain the Key Point matching at 57.50 he didn't see the words "minimum Euclidean Distance". That would have helped him a lot. It happens sometimes.
clear explanation .....thanks
does anyone have the matlab code for SHIFT?
Super clear, saved my ass.
Great tell. Can anyone tell me how to find or create datasets for detecting sexually explicit images?
Thank you. It is interesting video
well explained..
this is a UG course??
Yes it is taught in 3rd year ,but in fact in some places taught in 2nd year ,
great lecture thanks!!
Thankyou sir
thank you... :) this is awesome...
thanks a lot sir!
ONE WORD
AMAZING
Thanks for the bonus lecture in the end on How Google Search Works?
Thank you!
Dr. Shah is more concerned about citations than SIFT. I wish he described SIFT as detailed as the importance of citations (in his universe). This way there are still unexplained things.
thank so much
Nice!
thanks a lot very clear
Ok
nice though I didn't understand a bit !!
hahaha
Explain
i am doing same project on my PG
+Imam Vali hello
I'm working the same algorithm but I have problem to classify the result
can you help me
thanks a lot
Dear Imam Vali
At 26.30 time, What does it mean by every other rows and every other column in down sampling process?
cool :)
Damn it, Windows 7, get out of the way...
sift sucks surf rules!!!!
surf?
Dan Frederiksen google?
good explanation, thanks!
thank you for this