search

open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark
https://mmdetection.readthedocs.io
Apache License 2.0
28.71k stars 9.32k forks source link

Grounding DINO text prompts position exchange error #11033

Open luckyluckyjl opened 9 months ago

luckyluckyjl commented 9 months ago

Thank you very much for reproducing the training code for grounding dino. When I was using it, I found that after fine-tuning, the position exchange of prompt words can cause errors in the detection results.

hhaAndroid commented 9 months ago

@luckyluckyjl Of course, because you have a specific fine-tuning order.

luckyluckyjl commented 9 months ago

But if I reason according to the order of categories during fine-tuning, the result of the last category is still wrong. For example, in prompts: sedan . bus . bicycle . truck . excavator . Concrete truck . hazardous chemical truck. The detection results of last category is always wrong.

onnx20 commented 4 months ago

@hhaAndroid 你好，我遇到有类似的问题，不过我没有微调模型，直接使用你们预训练好的这个模型(MM-GDINO-L*/Swin-L/-/60.3/O365V2,OpenImageV6,ALL)，我发现使用相同的提示词，只是每次使用的类别顺序不一样，就有不一样的结果，甚至顺序变换后有些目标就检测不出来了，这个是有什么讲究么(比如：car . road . lake . truck . 交换顺序：road . lake . car . truck .)