Inconsistency during inference

[MLRadfys]

Hi,

I noticed that there are some inconsistencies when running inference on a single class in comparison with multiples classes. I.ex., I have an image containing persons and cars. When I run GroundingDINO on only the person class, all persons are detected nicely. When I run the model on both car and person class, some of the classes are detected as persons.

Is there any explanation for this behavior?

Thanks in advance,

M

[clementapa]

Same here

[rentainhe]

Hello, you may try the following methods for a better results:

• split the input classes with . like car. person.
• try this demo here by setting specific --token_span for car and person