results are different between roi pooling and roi align

[UcanSee]

Thanks for sharing such wonderful works！
I have used your codes to train OICR-vgg16， the mAP is 43% when I use roi-pooling. But when I use roi align, the mAP just have 23%. Difference between this two config is only 'POOLER_METHOD', Have you meets this problem in experiment?

[jason718]

#19 also reported this. I didn't use RoI-align in our exps, so this is also the first time I heard this. Let me take a look

[jason718]

maybe the issue is sampling_ratio?

in maskrcnn-benchmark, they set this to 2: https://github.com/facebookresearch/maskrcnn-benchmark/blob/57eec25b75144d9fb1a6857f32553e1574177daf/configs/pascal_voc/e2e_mask_rcnn_R_50_FPN_1x_cocostyle.yaml#L20

but this repo set the default value to 0:

wetectron/wetectron/config/defaults.py

Line 241 in e4fc028

_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0

So maybe try 2?

[UcanSee]

maybe the issue is sampling_ratio?

in maskrcnn-benchmark, they set this to 2: https://github.com/facebookresearch/maskrcnn-benchmark/blob/57eec25b75144d9fb1a6857f32553e1574177daf/configs/pascal_voc/e2e_mask_rcnn_R_50_FPN_1x_cocostyle.yaml#L20

but this repo set the default value to 0:

wetectron/wetectron/config/defaults.py

Line 241 in e4fc028

_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0

So maybe try 2?

I have tried sampling_ratio = 2， but the results have no difference.

[UcanSee]

Hello, have you conducted any experiments with FPN structure like R50-FPN? I never see any relational experiments in recent papers.

[jason718]

We did try, but it didn't outperform vanilla ResNet-C4 model. The multi-scale setting made it overfit to small regions somehow.

[UcanSee]

We did try, but it didn't outperform vanilla FPN model. The multi-scale setting made it overfit to small regions somehow.

what do you mean 'vanilla FPN model'? Does it mean that performance of MIST-R50-C4 model is better than MIST-R50-FPN model?

[jason718]

sorry typo, should be vanilla ResNet-C4 models. changed my prev answers

[UcanSee]

I've got it , thank you very much!