Encoding error by ultralytics prediction

[SamuelHafner]

Hi @mgonzs13,

thank you for your great work!!!

I have an error when when running the yolov8_node the self.yolo.predict cause an error (tried on 2 different cameras):

[yolov8_node-1] RuntimeError: Given groups=1, weight of size [48, 3, 3, 3], expected input[1, 4, 384, 640] to have 3 channels, but got 4 channels instead

I could fix this after i changed the desired encoding on the cv2_bridge:
from:

self.cv_bridge.imgmsg_to_cv2(msg)

to

self.cv_bridge.imgmsg_to_cv2(msg, desired_encoding='bgr8')

should we maybe add this permanent or is there some reason this encoding is empty?

[mgonzs13]

Hey @SamuelHafner, thanks for the issue. A better solution could be as follows. Can you try it?

cv_image = self.cv_bridge.imgmsg_to_cv2(msg, desired_encoding=msg.encoding)

[SamuelHafner]

Hey @mgonzs13, thanks for your fast response.
It is also not working because in my case I have an image with the encoding bgra8 and therefore an cv_image shape of (360, 640, 4). It has the alpha value too.

If I change the desired encoding for the image to bgr8 it converts the cv_image with the shape (360, 480, 3). And therefore this works for the yolo package
And I think that the yolo package can only handle images without alpha

[mgonzs13]

If you don't want to modify yolov8_ros, you can try to add a node to remove the alpha channel and republish the image as bgr8.

[SamuelHafner]

Yeah i could. But I added this desired_endcoding permanent now.
Would you like this change too? Should I do an merge request or just let it local.
Because normally you always would need an bgr8 encoding.

Btw. these recordings are from robocup@home 2023 right? We also were there. Which team? :)

[mgonzs13]

I'd rather create a new node to filter the alpha channel or modify the yolov8 node to remove it since there are cameras that do not use bgr8. Those bags were provided to me by the Gentlebots team.