Yolov5 inference image wrong colors

[unrue]

I'm running inference on Yolov5 trained net. Sometimes, some of the output image with bounding box has a strange blue dominant. Attach the original and the output. Have you ever experienced such behaviour? Thanks.

My code:

`for i in range(num_images):
    print("index: ", i, "image: ", inference_dir+'/'+img_list[i])
    im = cv2.imread(inference_dir+'/'+img_list[i])
    dimensions = im.shape
    results = model(im)
    results.save("img_"+str(i)+".jpg", save_dir='out_with_bb') 
    res_list.append(results.pandas().xyxy[0])
    img_dimensions_list.append(dimensions)
`

[jmarintur]

Hi @unrue, OpenCV uses BGR image format. You can simply do the following:

from cv2 import cvtColor, COLOR_BGR2RGB

for i in range(num_images):
    print("index: ", i, "image: ", inference_dir+'/'+img_list[i])
    im = cv2.imread(inference_dir+'/'+img_list[i])
    # get your image in RGB format
    im = cvtColor(im, COLOR_BGR2RGB)
    dimensions = im.shape
    results = model(im)
    results.save("img_"+str(i)+".jpg", save_dir='out_with_bb') 
    res_list.append(results.pandas().xyxy[0])
    img_dimensions_list.append(dimensions)

[unrue]

Thanks, but why It happens only on some images?

[jmarintur]

Mind sharing these images @unrue ?

[unrue]

The first one I already shared, the second one, is attached (it works with right colors)

[jmarintur]

Hi @unrue, I've just gone through the whole process, and I as mentioned before, you need to swap the bands. Here are the outputs I got using the images you shared:

Here is the code:

# image = "305950817-a0991c7d-e248-4125-842a-3abaf47ccd42.jpg"
image = "302577313-32bbc9a5-ca40-4331-abc9-66918effdbeb.jpg"
im = cv2.imread(image)
im = cvtColor(im, COLOR_BGR2RGB) # you could also swap manually the bands, i.e. #im = im[...,[2,1,0]]

results = model(im)
#results.save(image.replace('.jpg','_output.jpt'), save_dir='out_with_bb') 
results.show()

[unrue]

Ok, thanks!

[jmarintur]

@unrue, If it works for you, please remember to select the right answer. Thanks!

[unrue]

Actually the cluster where I did inference is offline I'll try as soon as possible.

[Mactarvish]

vis_image_np = src_image_np[..., ::-1] helps