Do you ever alter the form of the input image during the image inference process?

Support YOLO
, ,
1

ques
ques1565×525 5.12 KB

The color of this image was corrupted, so it was flagged by the detection system.

2

The color is not altered. It would only be wrong if you passed the wrong channel order. If you’re passing numpy array, the channel order should be BGR.

3

With Image.Open(path) as img:

image array = np.array(img.convert('RGB'))

arr = cv2.cvtColor(image array, cv2.COLOR_RGB2BGR)

model.predict(source=arr

,bath = 32

,verbose= False)

I performed inference on the image array after converting it, as shown in the source code above.

I passed the changed image array(RGB to BGR).

4

Your conversion path looks correct for a NumPy input. For model.predict(source=arr), NumPy arrays should be BGR, so RGB -> BGR before passing to Ultralytics YOLO is the right approach.

A simpler equivalent is either:

img = Image.open(path).convert("RGB")
results = model.predict(source=img, batch=32, verbose=False)

or:

arr = cv2.imread(path)  # already BGR
results = model.predict(source=arr, batch=32, verbose=False)

One small note: make sure the argument is batch, not bath.

YOLO preprocessing will resize/pad the image for inference, but it should not corrupt colors. If the “corrupted” color is from a saved/visualized image, check whether that viewer expects RGB while you are giving it BGR. That’s a common source of swapped colors. The Python prediction examples in the Ultralytics docs show both PIL and OpenCV/NumPy inputs.