RGB inputs or BGR inputs for model.predict(img)? about 6drepnet HOT 4 CLOSED

Hello,

the predict method converts the image into RGB. So it's okay, to pass a BGR array.
https://github.com/thohemp/6DRepNet/blob/master/sixdrepnet/regressor.py#L67

from 6drepnet.

Hello,

Thanks for your reply. I have already seen that in the code. The problem is the line above (https://github.com/thohemp/6DRepNet/blob/master/sixdrepnet/regressor.py#L66):

img = Image.fromarray(img)

that, I think, by default, converts an array with three channels into an RGB PIL image as it does not now which channel is what color. Thus img is already an RGB PIL image but with its R and B channels inverted. Convertimg it to an RGB PIL image should not invert its channels. Though I might be wrong. If I have some time today, I will add some print function around the code to check if I am correct or not

from 6drepnet.

You are correct, I just tested it. Swapping channels before using PIL should fix this:

img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

Thanks!

from 6drepnet.

Thanks for checking this and happy that helped fix a bug at the end!

from 6drepnet.