how should we annotate occluded object which is splitted in to two or more parts?

[xoofee]

This image is from coco
https://filebox.ece.vt.edu/~aroma/web/coco_images/val2014/COCO_val2014_000000005193.jpg

the original annotation of person is:

you can see that PERSON 1 annotation box does not contains his left palm, while PERSON 2 annotation box contains the left palm but not the right palm

does this lead to learning difficulty?

in fact the Yolov5x model's result on this image has a False Positive of Person which has a low IOU with the ground truth. The prediction does not contain PERSON 2's left hand. However, it contains the left hand of PERSON 1, which is wrong

python detect.py --weight yolov5x.pt --source coco-2017/data/000000005193.jpg

[glenn-jocher]

@xoofee annotations are correct, but of course this will be very difficult to predict correctly as these types of annotations represent a minority of the dataset.

[xoofee]

I got it! thanks!