search

ucbdrive / few-shot-object-detection

Implementations of few-shot object detection benchmarks
Apache License 2.0
1.08k stars 225 forks source link

COCO datasets: test the model you provide ,but the AP is only 5.002 #115

Closed xiaoliqiangxxqzz closed 2 years ago

xiaoliqiangxxqzz commented 3 years ago

When I used the 10shot model you trained on the Coco, I found that the AP was only 5.002 1620733393(1)

Is there anything I should pay attention to?

xiaoliqiangxxqzz commented 3 years ago

@thomasehuang @xinw1012 I'm sorry to bother you, but could you give me some advice?

scott-vsi commented 3 years ago

Which model/config are you running? That is very similar to the results reported, at least for the FRCN+ft-full model:

name nAP
FRCN+ft-full 5.5
TFA w/ fc 9.1
TFA w/ cos 9.1
xiaoliqiangxxqzz commented 3 years ago

Which model/config are you running? That is very similar to the results reported, at least for the FRCN+ft-full model:

name AP FRCN+ft-full 5.5 TFA w/ fc 9.1 TFA w/ cos 9.1

I use the model with cos classifier of 10 shot, and the models and data used are all provided by them.

xinw1012 commented 3 years ago

Hi, thanks for your interest in our work. Just to confirm, did you try to download the models from the model zoo directly? https://github.com/ucbdrive/few-shot-object-detection/blob/master/docs/MODEL_ZOO.md#10-shot-3? Can you verify the evaluation setups? We also released the training/eval logs in metrics so you may use it as an indicator for any discrepancy. There should not be a difference if you were using the pre-trained models. You can keep updating us if you have additional information on your experiment setups.

Thanks!