Has anyone run CBAM/BAM on fine grained dataset?

[goldentimecoolk]

I added CBAM and BAM in ResNet50 on fine grained dataset, e.g. CUB200. But it had a worse performance than pure ResNet50 backbone. Was it because I failed to have right training hyper parameters? Its also possible that the two methods are only effective for general classification but not suitable for fine grained classification. Dose anyone have the same experience or some suggestions? Thx!

[morning4346]

me too ,I trained it on cub.It is worse than Resnet50,I don't know why?

[goldentimecoolk]

me too ,I trained it on cub.It is worse than Resnet50,I don't know why?

It's worth to figure out the reason, which might largely result from the difference of general classification and fine grained classification. But it would be persuasive if we could analyse it by some quantitative methods.

[morning4346]

can you share to me about your code?perhaps something wrong in my code

[goldentimecoolk]

can you share to me about your code?perhaps something wrong in my code

I had the same result as you, i.e. a worse performance than Resnet50 backbone.
The main modifaction on your dataset is data loading and data augmentation part. Other parts are the same as source code.

[hezhu1996]

@goldentimecoolk @goldentimecoolk Hi, I also tested on object detection task in dataset Pascal voc, and it largely decrease the performance compared to the default resnet-50, have you solved this problem?

[cxy36]

@TWDH Hi, I had the same situation with you. On Pascal voc, the performance from Resnet-50+CBAM is worse than pure Resnet-50. I had checked my code again and again, and changed configs but doesn't work. Have you finally settled it?