FFMv1 for resnet-101 about m2det

[qijiezhao]

Hi, thanks for your interest in our paper.
We have modified the stride of Res4 from 2 to 1.
Then, the output of conv1, pool1, Res2, Res3, Res4, Res5 is:
1/2, 1/4, 1/4, 1/8, 1/8 and 1/16. We fuse the output of Res4 and Res5 as the base feature.

[qijiezhao]

So the stride from 2 to 1 doesn't damage the mAP?

Yes, you can directly use the ImageNet-pretrained model. Just like how Pspnet、deeplabv3 use the dilated Resnet.

[xinyuegtxy]

nice job, 但能把README训练部分写完整点吗？加上各种tricks的训练能开源下吗？thank you.

[qijiezhao]

@zealota @WangTianYuan Hi, thanks for your interest in reproducing our results. You can ask me any questions about it here.
By the way, I should warn you that, for training resnet101-M2det512 and VGG-M2det800, like the paper said, V100 GPU is better. Because of the limitation of batch size. In addition, you can tune the anchor scales to maximize the ability. For example, we have decreased the anchor's min size of the largest feature map(~25).

[WangTianYuan]

Do I need to retrain this modefied ResNet in ImageNet?

[zealota]

Do I need to retrain this modefied ResNet in ImageNet?

In my case, I use the pre-trained model on ImageNet.

[WangTianYuan]

So the stride from 2 to 1 doesn't damage the mAP?

[zealota]

@WangTianYuan, @qijiezhao
Sure, the mAP isn't damaged. I could achieve the 0.343 AP(IoU=0.50:0.95) on coco dataset, for the 320x320 input image.
And I used dilated ResNet.

[qijiezhao]

@WangTianYuan, @qijiezhao
Sure, the mAP isn't damaged. I could achieve the 0.343 AP(IoU=0.50:0.95) on coco dataset, for the 320x320 input image.
And I used dilated ResNet.

Congrats! By the way, dilated resnet is not a MUST, this step may slow down the inference speed.

[he-y]

Can you provide the pre-trained model for ResNet101.

[he-y]

@zealota
Hi, Can you provide the pre-trained model for ResNet101?
Thanks!