Can I directly train the model "FCOS_imprv_dcnv2_X_101_64x4d_FPN_2x" with giving config?

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)

https://arxiv.org/abs/1904.01355

Other

3.28k stars 630 forks source link

Can I directly train the model "FCOS_imprv_dcnv2_X_101_64x4d_FPN_2x" with giving config? #183

Closed Zeqiang closed 5 years ago

Zeqiang commented 5 years ago

GPU: 1080 Ti * 8 Config: fcos_imprv_X_101_64x4d_FPN_2x.yaml Can I directly train the model and get 46.6 mAP?

BTW, you mentioned in the page that: "State-of-the-art performance: Our best model based on ResNeXt-64x4d-101 and deformable convolutions achieves 49.0% in AP on COCO test-dev (with multi-scale testing)." How can I do the testing and get same mAP

And according to the paper, the best mAP is 44.7, what u have changed to pushed to 46.6?

tianzhi0549 commented 5 years ago

@Zeqiang All results can be reproduced by directly running our provided code. Please check out our model zoo for your other questions. https://github.com/tianzhi0549/FCOS#models

Zeqiang commented 5 years ago

@tianzhi0549 Thanks