Question about YoLoV4 vs EfficientDet

[liminghuiv]

I noticed that EfficientDetD0~D7 has image resolution of: 512, 640, 768, 896, 1024, 1280,1280, 1536.
YoloV4 has image resolution: 416, 512, 608.

@AlexeyAB , Does it mean that if YoLoV4 improve the image resolution and the corresponding settings, it can even have better performance than EfficientD7?

Thanks.

[WongKinYiu]

I think yes, we follow the same way as EfficientNet to optimize our anchors.

using optimized anchor for 416 to train:

• test on 320x320: 38.4 AP
• test on 416x416: 41.5 AP
• test on 512x512: 42.4 AP

using optimized anchor for 512 to train:

• test on 320x320: 37.7 AP
• test on 416x416: 41.2 AP
• test on 512x512: 43.0 AP
• test on 608x608: 43.5 AP

The image resolution and the corresponding settings can help improve AP.

By the way, EfficientDet is very powerful when inference on VPU or TPU...
See FPS information posted on #5079 (comment)

[liminghuiv]

Thanks. @WongKinYiu That's encouraging. I am interested in trying it. any suggestion on the settings?

[WongKinYiu]

Larger input resolution means larger objects to be detected. Currently YOLOv4 uses P3-P5 to detect objects. I think for input resolution 640-1024, we need P3-P6; for input resolution 1280-1536, we need P3-P7.

From the table, we can see EfficientDet gets better results on big objects when the input resolution increase.

[xevolesi]

Hi, @WongKinYiu !
Could you please give some insights about anchor-box generation methods or optimization methods?
I saw many issues where AlexeyAB said not to change standard anchor-boxes, but it's not clearly explained why shoudn't we change it?
Where can i read/watch/listen about anchor-box generation procedure and why i shoudn't change standard anchors? Could you please tell me? Thanks.

[AlexeyAB]

I saw many issues where AlexeyAB said not to change standard anchor-boxes, but it's not clearly explained why shoudn't we change it?

Because many people generate anchors but do not change masks, as described here: https://github.com/AlexeyAB/darknet#how-to-improve-object-detection

[xevolesi]

Hi, @AlexeyAB !
Thanks for clarification. I read these recommendations about anchor-boxes twelve times as you suggested in some of the issues. =)) But I still did not understand how do you notice, if the anchor-boxes bad or good? I saw issues where you saying that anchor-boxes are good/bad after seeing point cloud generated by kmeans procedure. If this related to

If many of the calculated anchors do not fit under the appropriate layers - then just try using all the default anchors.

or something else?

[beizhengren]

@AlexeyAB how to change the masks?