Output labels different from labels in input prompt.
#9
by
jcorsetti
- opened
Hello, after some experiments it seems that GroundingDino will output different labels from the ones provided in input. I tried with the following prompt: "s = "a chest of drawers. a door. a bed.", thus I expected GroundingDino to find only "chest of drawers", "door" or "bed". Instead one of the output labels is just "a chest". It seems that the first label I provided got truncated. Is this an expected behaviour?
Yes, when you use .post_processing_grounded_object_detection from GroundingDinoProcessor it uses the text_threshold to select the tokens. We could probably return both the original prompt and the thresholded prompt though, feel free to open an issue in the transformers repo