Class names and ID's seem to be flipped

[agillies8]

I have a yolov8 model which I trained with roboflow (exported the best.pt model weights after training in a colab notebook). The model has 3 classes: 0: blue box 1: green box 2: red box

I am able to run my model with this package, and everything seems to be working great. Im streaming images generated from Isaac Sim of 'boxes' going by on a conveyor. Im using this launch command inside the yolo_ros docker container:

ros2 launch yolo_bringup yolo.launch.py model:=src/models/best640.pt input_image_topic:=/camera/rgb

And here is the results, seen from rviz:

https://www.youtube.com/watch?v=1G3n6le86cw

However - as you can see, for some reason the classes seem to be inverted - red boxes are labeled as blue, and vice-versa, while green boxes are labeled correctly. When I test the model on roboflow, the classifications are correct. Any idea why this may be happening?

[agillies8]

Actually - i think this may have to do with inverted image and on RGB->BGR image. I wonder if i somehow trained on RGB images, but they are being fed as BGR images at runtime, since the color red and blue are defining features of the classes

[agillies8]

Ok, further investigation:

In your yolo.py file:

    def image_cb(self, msg: Image) -> None:

        if self.enable:

            # convert image + predict
            cv_image = self.cv_bridge.imgmsg_to_cv2(msg)
            results = self.yolo.predict(
                source=cv_image,
                verbose=False,
                stream=False,
                conf=self.threshold,
                iou=self.iou,
                imgsz=(self.imgsz_height, self.imgsz_width),
                half=self.half,
                max_det=self.max_det,
                augment=self.augment,
                agnostic_nms=self.agnostic_nms,
                retina_masks=self.retina_masks,
                device=self.device,
            )

CV2 might open images into BGR format, but then that gets passed right to yolo.predict, which expects rgb. See this:

[https://github.com/ultralytics/ultralytics/issues/9912]( )

import cv2
from ultralytics import YOLO

# Load the model
model = YOLO('yolov8n.pt')

# Read an image (in BGR format)
image_bgr = cv2.imread('path/to/image.jpg')

# Convert from BGR to RGB
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)

# Perform inference
results = model(image_rgb)

[agillies8]

yep, fixed it, that did the trick!

[mgonzs13]

@agillies8 thanks for the fix. I'll add the cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB) line to fix it.

[agillies8]

K, made a PR with the changes, here's it working now:

https://youtu.be/PtsvpXjY2wg