Is it possible to group coordinates per object type using "Point to <something>"?
Is it possible to group the coordinates per object type or do we need to run one inference step per object type?
I tried:
- "Point to object 1 then to object 2 etc"
- "Point to object 1. Point to object 2. etc"
But none of those works.
The model is able to point to multiple things at the same - can you share the exact prompt you tried & result you got?
"Point to all the black stones on the main go board. Point to all the white stones on the main go board."
I tried many different prompts. Can you share with me one that you know works?
Example attached
but maybe not what you meant by grouping
Sorry I should have been more clear. By grouping I mean instead of getting one single <points x1="8.9" y1="92.0" ></points>, I would get one for every objects:
<points x1="8.9" y1="92.0" alt="object1">object1</points><points x1="8.9" y1="92.0" alt="object2">object2</points>
The standard pointing mode does not support that, however you could try the point-question-answering mode. It can be a bit unreliable but might able to can handle that kind of request with the right prompt. To turn that on prefix you input query with: "point_qa:"
point_qa: seems to works well. Thank you!