Problems with results containing inaccurate words such as "bicyclebi", "bicyclebike", "bicycle motorbike", and "the".

IDEA-Research / GroundingDINO

[ECCV 2024] Official implementation of the paper "Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection"

https://arxiv.org/abs/2303.05499

Apache License 2.0

6.9k stars 698 forks source link

Problems with results containing inaccurate words such as "bicyclebi", "bicyclebike", "bicycle motorbike", and "the". #209

Open lanhas opened 1 year ago

lanhas commented 1 year ago

When the given text_prompt is “motorbike . bicycle . the telegraph pole ”, the result will contain “bicyclebi, telegraph pole, bicyclebike， telegraph, bicycle motorbike， the". I'd like to get some accurate categories that agree with the words I'm given. I know that raising the threshold or setting a condition to filter the recognition results would solve this problem, but it would result in a lot of missed boxes. How can I solve this problem?

lanhas commented 1 year ago

and those with #, such as '###bi'.

Dongjie-Cheng commented 9 months ago

Hi, I just met the same problem. Do you find a solution?

1benwu1 commented 6 days ago

this happens a lot... still don't know how to handle it