Normalized xywh values.

[SamiurRahman1]

❔Question

I am using crowdhuman dataset for training. I have my training annotations in xywh format but as far as i understand, in pixel values. So, i convert them as suggested in the instructions like this:

img = Image.open(root+image_name)
img = img.convert("RGB")
imgheight,imgwidth = img.size
x,y,w,h = a['hbox']
 yolo_x = x/imgwidth
 yolo_w = w/imgwidth
 yolo_y = y/imgheight
 yolo_h = h/imgheight

After training however, in the logs, the bboxes in example prediction and labels in images are misplaced. Is there anything wrong with my conversion method?

Additional context

[SamiurRahman1]

I am guessing i need to calculate the x_center like x_center = (x+w)/2 and then do: yolo_x = x_center/imgwidth is it correct?

[abidKiller]

i'm also facing the issue... maybe last update caused this bug

[SamiurRahman1]

I am guessing i need to calculate the x_center like x_center = (x+w)/2 and then do: yolo_x = x_center/imgwidth is it correct?

Tried this. did not work.

[glenn-jocher]

@SamiurRahman1 @abidKiller git clone the latest code and if you see errors follow the advice in https://docs.ultralytics.com/yolov5/tutorials/train_custom_data.

Errors in train and test jpgs indicate errors in your data.

[SamiurRahman1]

@SamiurRahman1 @abidKiller git clone the latest code and if you see errors follow the advice in https://docs.ultralytics.com/yolov5/tutorials/train_custom_data.

Errors in train and test jpgs indicate errors in your data.

this is the instruction that i followed. And about the error being in the data, the same annotations works perfectly when i draw a rectangle using the annotation data with cv2. but is my conversion method correct though?

[glenn-jocher]

@SamiurRahman1 conversion is explained in https://docs.ultralytics.com/yolov5/tutorials/train_custom_data

If your labels are correct your jpgs will be correct as well.

[abidKiller]

I used same dataset 2 days ago it was working fine, but when I tried to train today , i saw the misplaced bounding boxes in batch images.

[abidKiller]

As you can see , it was working perfectly

[glenn-jocher]

@abidKiller latest code is recommended.

[Xinlei-Ren]

@SamiurRahman1 @abidKiller git clone the latest code and if you see errors follow the advice in https://docs.ultralytics.com/yolov5/tutorials/train_custom_data.

Errors in train and test jpgs indicate errors in your data. Is it right to use the labelImg tool for labeling?

[glenn-jocher]

@Xinlei-Ren sure, you can use any labelling tool.

[SamiurRahman1]

Found the problem. As suspected, my conversion method was wrong. I tried training with a different dataset, it worked. So the only possibility was my conversion method. Here is how to convert the crowdhuman data:

imgheight,imgwidth = img.size
x,y,w,h = a['hbox']   //for each tag in gtboxes object
x_min = x 
x_max = (x+w)
y_min = y
y_max = (y+h)
x_center = (x_min+x_max)/2.
y_center = (y_min+y_max)/2.
yolo_x = x_center/imgheight
yolo_w = w/imgwidth
yolo_y = y_center/imgwidth
yolo_h = h/imgheight

hope this helps someone.

[glenn-jocher]

@SamiurRahman1 if you'd like and if you have time, to help others you can submit a PR for the addition of this dataset with i.e. data/crowdhuman.yaml, and a download script in data/scripts/crowdhuman.sh.

[nobody-cheng]

Found the problem. As suspected, my conversion method was wrong. I tried training with a different dataset, it worked. So the only possibility was my conversion method. Here is how to convert the crowdhuman data:

imgheight,imgwidth = img.size
x,y,w,h = a['hbox']   //for each tag in gtboxes object
x_min = x 
x_max = (x+w)
y_min = y
y_max = (y+h)
x_center = (x_min+x_max)/2.
y_center = (y_min+y_max)/2.
yolo_x = x_center/imgheight
yolo_w = w/imgwidth
yolo_y = y_center/imgwidth
yolo_h = h/imgheight

hope this helps someone.

why ??

yolo_x = x_center/imgheight
yolo_y = y_center/imgwidth
# =================
yolo_x = x_center/imgwidth
yolo_y = y_center/imgheight

[glenn-jocher]

@nobody-cheng My apologies for the confusion. You are absolutely right. The correct conversion should be:

yolo_x = x_center/imgwidth
yolo_w = w/imgwidth
yolo_y = y_center/imgheight
yolo_h = h/imgheight

Thank you for catching that mistake!