Ok, so the histograms are calculated per cell or per block? I'm missing something at this point... Because if they are calculated at cell level, then we use the blocks only to define cells?
How can we make a HOG descriptor out of binary image(black and white image) where the value just 1 or 0. Because in the paper that i am using right now it using the HOG method to make a feature vector out of a sketch binary image.
NO.. here our consideration is 85 degrees .. and its distance from 70 is 15 degrees which means its chances of getting added to that bin is less(hence 1/4 as per calc.) where as for 90 , distance is just 5 so its chances are more to enter that bin(hence 3/4 ).. Hope you got the point
There seems to be redundant data during concatenation. E.g. if you have cells 1 2 3 4 5 6 7 8 9 10 Block 1 contains 1,2,6,7 and Block 2 contains 2,3,7,8 and so on. And after concatinating the histograms, we will get a vector that is something like this H1,H2,H6,H7,H2,H3,H7,H8.... This histograms from cell 2 and 7 are repeated in the vector. Same will apply for many other cells. Technically, just concatenating the histograms of the cells should suffice? H1 H2 H3 H4 .... or any particular order Unless there is some hidden significance of this the feature vector seems to be longer than needed. Am I missing something?
"In fact, we typically overlap the blocks so that each scalar cell response contributes several components to the final descriptor vector, each normalized with respect to a different block. This may seem redundant but good normalization is critical and including overlap signicantly improves the performance." What the hell does "each normalized w.r.t. a different block" mean?
How does orientation binning work for 0-180 degrees? How does he account of the rest of the gradient orientations? I know that it might be signed -180 to 180, but how doesn't he make the histogram from -180 to 180 then?
Thank you but I have 2 questions for this issue that i dont understand clearly: 2. After using HOG, we will get a feature vector for input image matrix, right? 2. I think we must prove this method by matrix, i mean input matrix, so with HOG, where is matrix of image?
![Lecture 9.3: Features [Histogram of Gradients] [HOG]](http://i.ytimg.com/vi/92-NaIdIrEs/mqdefault.jpg)
I checked this video after reading the Dalal and Triggs paper and everything became clearer. Great post. Thanks a lot.
How are the HOG feature vectors visualized at 10:08? what exactly is being plotted to visualize the features?
Thank you so much professor, I am looking for the answer for the re sample image for using HOG and found it. 14:05
I have a question . Image has 105 blocks and each block has 9 bins. But the features dimension is 105*4*9 so what does 4 stand for ?
each block is 2*2 cells(4) , each cell has 9 bins...
Sai Bhargav Avula oh thanks so much. it's clear for me now
Ok, so the histograms are calculated per cell or per block? I'm missing something at this point... Because if they are calculated at cell level, then we use the blocks only to define cells?
In each cell , the histogram will be
calculated.
@@saibhargavavula8881 but no. of cells =16*8=128
And each cell has 9bins. So shouldn't it be 128*9=1152
Thank you, professor, I was also searching how that HOG was implemented, and now the idea is clear. Thank you very much, sir.
I guess it's kinda off topic but do anyone know of a good website to watch newly released series online?
@Anders Zion I dunno lately I've been using Flixportal. just search on google after it =) -jameson
@Jameson Jaden Thank you, signed up and it seems like they got a lot of movies there :) Appreciate it!!
@Anders Zion You are welcome =)
How can we make a HOG descriptor out of binary image(black and white image) where the value just 1 or 0. Because in the paper that i am using right now it using the HOG method to make a feature vector out of a sketch binary image.
so then, is the unit of matlab output of extractHOGfeatures command of in radian? if this descriptor spits out dominant direction, right?
After the normalization of the blocks what happens to the four different histograms of the same cell?
Minute 6:33, why did he exchange the ratios? 1/4 Should be related to Bin 90 and 3/4 should be related to Bin 70. Right?
NO.. here our consideration is 85 degrees .. and its distance from 70 is 15 degrees which means its chances of getting added to that bin is less(hence 1/4 as per calc.) where as for 90 , distance is just 5 so its chances are more to enter that bin(hence 3/4 ).. Hope you got the point
There seems to be redundant data during concatenation. E.g. if you have cells
1 2 3 4 5
6 7 8 9 10
Block 1 contains 1,2,6,7 and Block 2 contains 2,3,7,8 and so on. And after concatinating the histograms, we will get a vector that is something like this
H1,H2,H6,H7,H2,H3,H7,H8.... This histograms from cell 2 and 7 are repeated in the vector. Same will apply for many other cells. Technically, just concatenating the histograms of the cells should suffice? H1 H2 H3 H4 .... or any particular order
Unless there is some hidden significance of this the feature vector seems to be longer than needed. Am I missing something?
Amol Borkar The trick is in the normalization
"In fact, we typically overlap the blocks so that each scalar cell response contributes several components to the final descriptor vector, each normalized with respect to a different block. This may seem redundant but good normalization is critical and including overlap signicantly improves the performance." What the hell does "each normalized w.r.t. a different block" mean?
how does the numtiple human detection in a single frame using hog ?
Can you tell how to analysis the performance ratio?
How does orientation binning work for 0-180 degrees? How does he account of the rest of the gradient orientations? I know that it might be signed -180 to 180, but how doesn't he make the histogram from -180 to 180 then?
The bins are made unsigned since (I guess) the HOG is basically an histogram distribution over EDGEs and edges are sign insensitive.
who can give me a numerical examples about HOG descriptor? I need it
Excellent Description
Thanu you Dr. Shah
Thank you sir. you delivered it well #UserFromPakistan
does anyone know what's the deficiency of HOG feature
great description thanks
Thank you but I have 2 questions for this issue that i dont understand clearly:
2. After using HOG, we will get a feature vector for input image matrix, right?
2. I think we must prove this method by matrix, i mean input matrix, so with HOG, where is matrix of image?
Happy holi and dhuladi
Why not illustrate the hog calculation with one exemple, in order to see what is magnitude vote etc.... It is too much bla bla !!!
it is cropped
Got clear as compare to SIFT
#boss
thanks professor
Thank you sir..
Great
Yes, we are indian