{ "id": 7342, "title": "Importance of colours of images in predictions", "url": "/competitions/galaxy-zoo-the-galaxy-challenge/discussion/7342", "author_name": "", "post_date": "2014-03-10T16:36:24.343Z", "votes": null, "comment_count": 8, "views": 2731, "content": "

I did small test to see if there is difference between using only gray channel features (features computed from rgb->gray converted images) vs RGB features. I would think that converting RGB to HSV might provide better insights, but I did not try that yet.

\n

I found out (using gradient boosting variable importance) that many Green and Blue features come out as first in variable importance before any gray channel only features. This I just tested for class1.1, class 1.2, class1.3 but I suppose similar findings can be done for them.

\n

Also I noticed that (some) green channel features are significantly more important than blue or red channel features [again checked only for 1.1,1.2 and 1.3 class outputs]. This might be related to fact that human eye observes blue wavelength worse than green or red. See for example http://www.normankoren.com/Human_spectral_sensitivity_small.jpg

\n

(that is why I used originally heavy downweighting of blue channel when converting RGB images to gray scale).

\n

 Anyone else tested importance of RGB channels (and willing to share any information :))?

", "messages": [ { "id": "40103", "postDate": "03/10/2014 16:36:24", "content": "

I did small test to see if there is difference between using only gray channel features (features computed from rgb->gray converted images) vs RGB features. I would think that converting RGB to HSV might provide better insights, but I did not try that yet.

\n

I found out (using gradient boosting variable importance) that many Green and Blue features come out as first in variable importance before any gray channel only features. This I just tested for class1.1, class 1.2, class1.3 but I suppose similar findings can be done for them.

\n

Also I noticed that (some) green channel features are significantly more important than blue or red channel features [again checked only for 1.1,1.2 and 1.3 class outputs]. This might be related to fact that human eye observes blue wavelength worse than green or red. See for example http://www.normankoren.com/Human_spectral_sensitivity_small.jpg

\n

(that is why I used originally heavy downweighting of blue channel when converting RGB images to gray scale).

\n

 Anyone else tested importance of RGB channels (and willing to share any information :))?

", "rawMarkdown": "", "votes": null }, { "id": "40161", "postDate": "03/11/2014 12:51:53", "content": "

Hello,

\n

I did no detailed measurements but this is what I concluded:

\n

Color was better than GrayScale. I run a number of experiments and the difference was noticable.

\n

I also tried HSV but did not see any improvements so I went back to RGB.

", "rawMarkdown": "", "votes": null }, { "id": "40177", "postDate": "03/11/2014 19:09:40", "content": "

As Jean Tate discussed in this thread, colour is quite informative: http://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge/forums/t/6730/do-we-require-color-images-or-can-we-work-on-grayscaled

\n

If it wasn't, the benchmark wouldn't make much sense anyway :)

", "rawMarkdown": "", "votes": null }, { "id": "40182", "postDate": "03/11/2014 20:23:31", "content": "

A nice, simple color feature is mentioned in the Benjari et al. paper. In this paper it is called the (g-r) and (r-i) colour.

\n

This is simply the log of the relative color density. I use for instance -10*log(total_green_colour / total_red_colour) and the same for red over blue. It becomes more informative if you can extract these from just the galaxy (as opposed to the whole picture).

", "rawMarkdown": "", "votes": null }, { "id": "40185", "postDate": "03/11/2014 21:46:45", "content": "

[quote=sedielem;40177]

\n

If it wasn't, the benchmark wouldn't make much sense anyway :)

\n

[/quote]

\n

The benchmark really doesn't make any sense. The RMSE obtained with it is only slightly (very slightly) better then the RMSE of predicting the average label over the training data (which is 0.16374).

\n

a 0.001 improvement might be statistically significant with such a large data set, but it can also be explained with the central pixel intensity value and not necessarily it's color (I didn't really look at the implementation but I assume that the intensity wasn't disregarded on purpose)...

my guess is that the color itself is not really important directly, but rather it's easier for humans to see color and intensity differences over just intensity differences and that the main influence of the color info is this improved human discrimination between shapes. central bulge prominent or not? dense spiral or regular elliptical? stuff like that.

but anyway, enough theorizing, back to coding.

", "rawMarkdown": "", "votes": null }, { "id": "40187", "postDate": "03/11/2014 22:11:42", "content": "

Fair point about the benchmark ;)

\n

With regards to the second part of your post, as Jean Tate said in the other thread I linked, there is more to it than just improved human discrimination: "spirals and irregular galaxies are bluer than ellipticals".

", "rawMarkdown": "", "votes": null }, { "id": "40214", "postDate": "03/12/2014 13:37:19", "content": "

Well, in that case I guess that even if were to be right and as it happens humans themselves don't use the color info directly to distinguish spiral from elliptical galaxies (I was thinking that they don't have the time or need to learn any more complex color relationships with the 50 classification they have to make throughout the experiment), then it doesn't really matter in practice because machine learning algorithms will surely be able to use this consistent color info to compensate a little bit for their inherent inferiority with tasks like understanding the shape of an object.

", "rawMarkdown": "", "votes": null }, { "id": "40243", "postDate": "03/13/2014 05:25:25", "content": "

I tried adding some features which were colour-dependent (mostly they were the same as features for the grey image, except applied to each of the channels). All of them improved the result, some by not very much. Checking out the feature importances after training the classifiers seems to indicate that the green channel has the less value, while the red and blue have the most (seems to be consistent with irregular -> bluer than elliptical).

", "rawMarkdown": "", "votes": null }, { "id": "2327748", "postDate": "07/03/2023 06:32:39", "content": "

Reatime RGB vs HSV ans Lab Channels : https://youtu.be/5dJcuGmH7K0

", "rawMarkdown": "Reatime RGB vs HSV ans Lab Channels : https://youtu.be/5dJcuGmH7K0", "votes": null } ], "comments": [ { "id": 2327748, "author_name": "bemorekgg", "author_url": "", "post_date": "07/03/2023 06:32:39", "content": "

Reatime RGB vs HSV ans Lab Channels : https://youtu.be/5dJcuGmH7K0

", "votes": null, "replies": [] }, { "id": 40161, "author_name": "juliandewit", "author_url": "", "post_date": "03/11/2014 12:51:53", "content": "

Hello,

\n

I did no detailed measurements but this is what I concluded:

\n

Color was better than GrayScale. I run a number of experiments and the difference was noticable.

\n

I also tried HSV but did not see any improvements so I went back to RGB.

", "votes": null, "replies": [] }, { "id": 40177, "author_name": "sedielem", "author_url": "", "post_date": "03/11/2014 19:09:40", "content": "

As Jean Tate discussed in this thread, colour is quite informative: http://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge/forums/t/6730/do-we-require-color-images-or-can-we-work-on-grayscaled

\n

If it wasn't, the benchmark wouldn't make much sense anyway :)

", "votes": null, "replies": [] }, { "id": 40182, "author_name": "osmits", "author_url": "", "post_date": "03/11/2014 20:23:31", "content": "

A nice, simple color feature is mentioned in the Benjari et al. paper. In this paper it is called the (g-r) and (r-i) colour.

\n

This is simply the log of the relative color density. I use for instance -10*log(total_green_colour / total_red_colour) and the same for red over blue. It becomes more informative if you can extract these from just the galaxy (as opposed to the whole picture).

", "votes": null, "replies": [] }, { "id": 40185, "author_name": "selfishgene", "author_url": "", "post_date": "03/11/2014 21:46:45", "content": "

[quote=sedielem;40177]

\n

If it wasn't, the benchmark wouldn't make much sense anyway :)

\n

[/quote]

\n

The benchmark really doesn't make any sense. The RMSE obtained with it is only slightly (very slightly) better then the RMSE of predicting the average label over the training data (which is 0.16374).

\n

a 0.001 improvement might be statistically significant with such a large data set, but it can also be explained with the central pixel intensity value and not necessarily it's color (I didn't really look at the implementation but I assume that the intensity wasn't disregarded on purpose)...

my guess is that the color itself is not really important directly, but rather it's easier for humans to see color and intensity differences over just intensity differences and that the main influence of the color info is this improved human discrimination between shapes. central bulge prominent or not? dense spiral or regular elliptical? stuff like that.

but anyway, enough theorizing, back to coding.

", "votes": null, "replies": [] }, { "id": 40187, "author_name": "sedielem", "author_url": "", "post_date": "03/11/2014 22:11:42", "content": "

Fair point about the benchmark ;)

\n

With regards to the second part of your post, as Jean Tate said in the other thread I linked, there is more to it than just improved human discrimination: "spirals and irregular galaxies are bluer than ellipticals".

", "votes": null, "replies": [] }, { "id": 40214, "author_name": "selfishgene", "author_url": "", "post_date": "03/12/2014 13:37:19", "content": "

Well, in that case I guess that even if were to be right and as it happens humans themselves don't use the color info directly to distinguish spiral from elliptical galaxies (I was thinking that they don't have the time or need to learn any more complex color relationships with the 50 classification they have to make throughout the experiment), then it doesn't really matter in practice because machine learning algorithms will surely be able to use this consistent color info to compensate a little bit for their inherent inferiority with tasks like understanding the shape of an object.

", "votes": null, "replies": [] }, { "id": 40243, "author_name": "kunstmord", "author_url": "", "post_date": "03/13/2014 05:25:25", "content": "

I tried adding some features which were colour-dependent (mostly they were the same as features for the grey image, except applied to each of the channels). All of them improved the result, some by not very much. Checking out the feature importances after training the classifiers seems to indicate that the green channel has the less value, while the red and blue have the most (seems to be consistent with irregular -> bluer than elliptical).

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "40103": "", "40161": "", "40177": "", "40182": "", "40185": "", "40187": "", "40214": "", "40243": "", "2327748": "Reatime RGB vs HSV ans Lab Channels : https://youtu.be/5dJcuGmH7K0" }, "source": "meta" }