SSDlite annotation format

Hello, I am fine-tuning pretrained SSDlite model from torchvision.
The training somehow goes well when I feed bounding boxes coordinates in absolute coordinates to the fed image size. However, the bb coordinates “theoretically” should be normalized.
I’d like to know that In what format does the SSDlite in torchvision expect the annotation?
In addition to it, what is the reason that the absolute coordinates make training going well while relative coordinates do not.

just in case, absolute coordinates look like this (162, 280, 185, 300). relative coordinates look like this (0.52, 0.8, 0.6, 0.9).

I don’t know how exactly you are feeding the bounding box coordinates to the training, but based on the model definition the output coordinates don’t seem to be normalized:

        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.

thank you for the answer. I could confirm the reason.
by the way, the classification loss is not decreasing very well after making annotations to absolute coordinates. Is this because of loss functions in SSDlite or something else?