search

IDEA-Research / Grounded-Segment-Anything

Grounded SAM: Marrying Grounding DINO with Segment Anything & Stable Diffusion & Recognize Anything - Automatically Detect , Segment and Generate Anything
https://arxiv.org/abs/2401.14159
Apache License 2.0
15.12k stars 1.4k forks source link

Phrases not matching the Prompt given #240

Open sachidanandvs opened 1 year ago

sachidanandvs commented 1 year ago

Hi, I am trying to evaluate Grounded SAM on COCO instance segmentation dataset. For that, I am giving a text prompt which is a sentence with all coco classes separated by a comma, and the Grounding dino gives bounding boxes and its corresponding class as a phrase, but sometimes these phases are not matching the given prompt, some times it doesn't even mean anything. Sometimes I am getting multiple classes for a single bounding box, which are not supercategories of each other. Like the phrases are stemmed words of a given text prompt, like fire hydrant is given as firehydr, sometimes it is outside of class like, I got '##hy' as a pred_phrase for a bounding box which is not in the text prompt. Is there any code given for mapping pred_phrases to its corresponding class given in the text prompt.

Glisten-5481 commented 1 year ago

I have encountered the same problem and hope to receive help.

meidachen commented 1 year ago

Any help on this issue would be appreciated!