search

WXinlong / DenseCL

Dense Contrastive Learning (DenseCL) for self-supervised representation learning, CVPR 2021 Oral.
GNU General Public License v3.0
549 stars 70 forks source link

The performance of detection in COCO #7

Closed Peipeilvcm closed 3 years ago

Peipeilvcm commented 3 years ago

Based on MMDetection，train COCO2017 & val COCO2017

FasterR-CNN,r50 From torchvision://resnet50

       1x: bbox_mAP: 0.3750

FasterR-CNN,r50 From My Reproduction Model Pretrained on ImageNet

       1x: bbox_mAP: 0.3580

FasterR-CNN,r50 From Your Pretrained Mode on ImageNetl

       1x: bbox_mAP: 0.3550

Which is not as good as expected? Could you give a help？

WXinlong commented 3 years ago

Did you use SyncBN for each component during the detector training? If not, you should do so. Please see this config for reference. The reason can be found in MoCo paper.

Peipeilvcm commented 3 years ago

Did you use SyncBN for each component during the detector training? If not, you should do so. Please see this config for reference. The reason can be found in MoCo paper.

Thanks for your quick reply, I will try it, and updates the results

Peipeilvcm commented 3 years ago

Based on MMDetection，train COCO2017 & val COCO2017, No freeze stage,Add SyncBN

FasterR-CNN,r50 From torchvision://resnet50

       1x: bbox_mAP: 0.3740

FasterR-CNN,r50 From My Reproduction Model Pretrained on ImageNet

       1x: bbox_mAP: 0.3780

FasterR-CNN,r50 From Your Pretrained Mode on ImageNetl

       1x: bbox_mAP: 0.3750