search

ucbdrive / few-shot-object-detection

Implementations of few-shot object detection benchmarks
Apache License 2.0
1.12k stars 225 forks source link

can not detect novel classes #171

Open ysw774977 opened 2 years ago

ysw774977 commented 2 years ago

There are two classes in my dataset. I modified the number of classes in the yaml file of base.yaml to 1, the number of classes in the novel.yaml to 1, and the number of classes in the all.yaml to 2. However, the final model can only predict the classes in base, but the classes in novel cannot.The AP is high.The AP of base class is very high, but the AP of novel class is 0.

salehnia commented 2 years ago

@ysw774977 @xinw1012 I have same problem (just in results), The AP of base class is good, but in novel class it's bad (0.09 in 5shot). When I checked log file, I saw it's because of "loss_rpn_cls" value, It's almost fix on 0.2 during training. I don't know how I can improve it, I changed LR value, but it wasn't effective.

ysw774977 commented 2 years ago

@ysw774977 @xinw1012 I have same problem (just in results), The AP of base class is good, but in novel class it's bad (0.09 in 5shot). When I checked log file, I saw it's because of "loss_rpn_cls" value, It's almost fix on 0.2 during training. I don't know how I can improve it, I changed LR value, but it wasn't effective.

The reason for this problem is often that the scale difference between the two is too large. I think when the scale difference between the detected objects is too large, it is not suitable to fine tune with TFA method, because it is too difficult for RPN.

salehnia commented 2 years ago

@ysw774977 thank you for your reply.