According to your paper, objects that smaller than 2% will be ignored. https://github.com/google/sg2im/blob/26383737500a2855e772637a222bb0a871034f0d/sg2im/data/coco.py#L152 but this code only consider the width and height, and this could not reflect the true area of an object, For example, this is a "clothes" stuff, and most of the area is blank. (Original image: http://cocodataset.org/#explore?id=245764)  the area calculated by w and h is 10692, but if you use codes like `np.sum(mask)', you get a more accurate area of 1115.
According to your paper, objects that smaller than 2% will be ignored.
sg2im/sg2im/data/coco.py
Line 152 in 2638373
but this code only consider the width and height,
and this could not reflect the true area of an object,
For example, this is a "clothes" stuff, and most of the area is blank. (Original image: http://cocodataset.org/#explore?id=245764)
the area calculated by w and h is 10692,
but if you use codes like `np.sum(mask)', you get a more accurate area of 1115.