Hessian-Laplace blob detector.

[Daniil Pakhomov]

Hello,

I want to try to implement Hessian-Laplace blob detector (as mentioned in requested features on github page).

Can someone give me the list of corresponding papers, using which I can implement it.

Thank you. 

[Daniil Pakhomov]

Hello,

I would like to implement this detector.

Can someone give me a list of papers that may help.

Thank you.

[Vighnesh Birodkar]

Hello

The Hessian Laplace blob detector was explained in [1].
Also, look at see `skimage.feature.blob_doh` and skimage.feature.blob_log`

[1] : http://www.robots.ox.ac.uk/~vgg/research/affine/det_eval_files/mikolajczyk_ijcv2004.pdf

Thanks
Vighnesh

[Daniil Pakhomov]

Hello, Vighnesh Birodkar.

Thank you for your reply.

I have read through some of the related papers and I have a question:

As it seems to me: I will have to compute Hessian with different sigmas (to find probable feature points) and also Laplacian of Gaussian with different sigmas( to check whether or not the points found on the previous step are local minimas in the scale space).

What do you think about efficiency? Because if I have to compute two image cubes, it will take some time.

вторник, 3 марта 2015 г., 6:18:22 UTC+1 пользователь Vighnesh Birodkar написал:

[Daniil Pakhomov]

Now I have a more well-formed question:

Do you think it is also feasible to approximate laplacian of gaussian with haar wavelets?

понедельник, 2 марта 2015 г., 18:05:09 UTC+1 пользователь Daniil Pakhomov написал:

[Vighnesh Birodkar]

Hello Daniil

Unfortunately, I am not that well-informed to comment about haar wavelets, I will definitely read up though.

Constructing 2 image cubes is not required, you will construct one to determine (x,y) coordinates of maximas using the the determinant of hessian image cube.
Once that is done, for only those (x, y) points you will compute Laplacian Of Gaussian for different scales, and final the scale space maxima for only those points.

This gives us the best of both approaches.

Thanks
Vighnesh

[Daniil Pakhomov]

Great! Thank you.

What I am thinking about is to take your _hessian_matrix_det() and make it also return d_xx + d_yy for each element. So, on the output I will get determinant of Hessian and also a Laplacian.

I have a small question: d_xx + d_yy will be a scale normalized Laplacian in your notation? 

понедельник, 2 марта 2015 г., 18:05:09 UTC+1 пользователь Daniil Pakhomov написал:

Hello,

[Daniil Pakhomov]

Really sorry for spamming you with questions.

No more need to answer.

I implemented this detector and it works as fast as your determinant of Hessian approach implementation.

It passes all you tests and works better with coin() images (it doesn't detect a false coin as determinant of Hessian does in the example).

May I ask you to do a review of my code later?

Thank you.

понедельник, 2 марта 2015 г., 18:05:09 UTC+1 пользователь Daniil Pakhomov написал:

Hello,

[Josh Warner]

We'd welcome this as a PR on GitHub. That would be the ideal place for code review, etc.

[Даниил Пахомов]

Thanks.

I've sent it.

--
You received this message because you are subscribed to a topic in the Google Groups "scikit-image" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/scikit-image/ghIYwQFubEU/unsubscribe.
To unsubscribe from this group and all its topics, send an email to scikit-image...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Vighnesh Birodkar]

Hey Daniil

A good technical discussion is always welcome. Its never considered as spamming.

Thanks
Vighnesh