Hold up... Are these tensorflow models trained in RGB or BGR?

Question

Hold up... Are these tensorflow models trained in RGB or BGR?

apockill opened this issue 7 years ago · 5 comments

I have reason to believe that the tensorflow models have been trained with image arrays in RGB format, but this codebase loads images using OpenCV, which defaults to BGR.

This issue is plaguing me at work on a different project, and I want to confirm: Should I feed it an RGB image or a BGR image?

If anyone knows, this would be quite helpful!

Answer 1 · 2017-10-03T20:11:42.000Z

hi @apockill

no idea, but if it's the case, I wonder if that would be what causes no detection in the app, but great detection using the object_detection jupyter notebook.

datitran/raccoon_dataset#25

Answer 2 · 2017-10-03T23:43:07.000Z

Hahaha oh man I'm so glad I found this issue. Yeah, that's the problem. Throw a cv2.cvtcolor(bgr2gray) or whatever in there, should run great!

Answer 3 · 2017-10-04T13:14:51.000Z

@apockill

3 questions

doesn't that turn your video gray?
where did you put it?
what do you mean by "or whatever"? are you saying that it's bgr and needs to be changed to anything but bgr?

Answer 4 · 2017-10-04T15:42:21.000Z

I must have been really tired/lazy when I typed that out. I meant cv2.cvtColor(img, cv2.COLOR_BGR2RGB). That worked for me, and increased performance of my networks. I change the image format before feeding the input into the network.

…

On Wed, Oct 4, 2017, 6:14 AM lhammond ***@***.***> wrote: @apockill <https://github.com/apockill> 3 questions 1. doesn't that turn your video gray? 2. where did you put it? 3. what do you mean by "or whatever"? are you saying that it's bgr and needs to be changed to anything but bgr? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#29 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABqOfIHvlWefB1MlEwTaOqinMn2W3Cvlks5so4TNgaJpZM4PpqcY> .

Answer 5 · 2017-10-04T20:14:18.000Z

Hey @apockill, thanks for pointing this out. You're right. Actually for another project I was doing the bgr2rgb and vice versa conversion and the raccoon video is also used with this codebase. I updated the code. Next time you could also do a pull request! Thanks again!