What change it up to 53.4? And vs DetectoRS.

[zcode86]

I see that the mAP has increased to 53.4 again, I would like to ask what has changed to cause it up from 52.6? And is it stronger than DetectoRS? Because the DetectoRS used multi-scale, while efficientdet used single scale, and FPS is also bigger.

[mingxingtan]

We fixed some dataloader issues (thanks to a lot of contributors), and I retrained it with longer time.

What is the AP and FPS for DetectoRS?

[zcode86]

Thanks for your reply. On COCO test-dev, DetectoRS achieved 54.7 when using ResNeXt-101-32x4d as backbone and TTA(test-time augmentation,which includes multi-scale testing, horizontal flipping,etc.), while 53.3 no TTA. And they only show FPS=3.9 when using ResNet-50 as backbone and no TTA, which achieved 51.3. So I guess it will cost much time if want to achieve higher mAP.

[mingxingtan]

Thanks for the info @iamastar88

I noticed this paper is using soft-nms, so I also added soft-nms yesterday. Using the same ckpt, softnms (gaussian with sigma=0.5) improves D7 COCO AP from 53.4 to 53.7

[zcode86]

Wow, that's amazing.