Grounded-Segment-Anything
Grounded-Segment-Anything copied to clipboard
IDEA-Research
multiple object inference
Hi, thanks for the great implementations. I am trying to run this code with the multi-object prompt. As you guided, I used ' . ' between each classes, but could get a bit weird results. I have two questions with my one sample output image.
- My prompt is "asparagus . carrot . zucchini ." Is there some mistakes with this prompt? I am wondering why the model predicts all the asparagus as zucchini, and I don't think this is only because of the DINO performance issue.
- Is there a case that we can see broken class output such as '##ni' on the image?
Thanks again!
Same question here. I tried "cat . dog ." on asset/demo6.jpg, but I got wierd output phrase "cat dog". I also tried "cat,dog", it does return cat or dog. But the number of detected objects decreases and some objects are misclassified.
Has anyone solved the issue? Not yet solved from my side yet.
I also face this problem.
Another question for me: I can get correct segmentation with single prompt (e.g. cat ), however I cannot segment out this object with multi-object prompt (e.g. like cat . dog .)
Have you noticed this problem?
@longyangqi yeah, that issue is just what I am facing recently. Hope the authors are taking care of it.
@mhyeonsoo The "##ni" is caused by the tokenizer (something like splitting zucchini into "zucchi" + "##ni") you can have a look at https://github.com/IDEA-Research/Grounded-Segment-Anything/blob/main/GroundingDINO/groundingdino/util/inference.py