New evidence format#9
Conversation
|
I don't think we can use pixels for coordinates. As PDFs have no native resolution, document.json should be resolution-agnostic. I suggest using relative positions (normalized by the height and width). |
|
Ok, we have bounding boxes in pixels returned by the model and in some datasets, in both cases normalised values are used and bounding box in pixels is just additional optional information, I'm going to consult on solutions and notify here. |
|
We could use bounding boxes in pixels internally and, here, use just the normalized values. |
|
Ok, I removed bounding boxes in pixels from the schema. |
|
I changed fields "positions" and "group_item_ids" in evidence to optional. |
|
It looks good now, but please add an example of how to use group_item_ids in the examples directory. |
|
group_item_ids is a field used only by graphical interface in our services to know how to highlight bounding boxes, in our work we set it always to 0 currently. |
|
Is it ok to leave group_item_ids set to zeros in examples? |
Updating schema for new evidence format