Any plan to support pp-yolo?

[toplinuxsir]

Any plan to support pp-yolo? Thanks!
https://arxiv.org/abs/2007.12099

[AlexeyAB]

• SPP +0.3% AP (marked as ”star”)

• Coord Conv +0.5% AP (marked as "diamonds" for each 1x1-conv) https://arxiv.org/abs/1807.03247

• IoU Aware +0.6% AP (predicted_IoU === bbox_confidence, a bit like Gaussian_yolo) https://arxiv.org/abs/1912.05992

• Matrix NMS +0.7% AP https://arxiv.org/abs/2003.10152

• DropBlock (marked as "triangles" in FPN) + EMA (Wema = λWema + (1 − λ)W, λ = 0.9998) + LB (mini-batch=196) = +2.3% AP

  • EMA: https://arxiv.org/abs/1911.09070

• DCN : https://arxiv.org/abs/1703.06211

  • we replace the 3×3 convolutional layer in the last stage of ResNet with deformable convolution layer

---

• DropBlock – should be used only FPN (1 DropBlock per 1 FPN-branch)
• DCN – should be used only for the conv-3x3 in the last stage of Backbone
• Coord Conv – should be used only for conv-1x1 in FPN / head
• We can add another one probability = IoU_prob. So prob = objectness * IoU_prob * Class_prob
• Matrix NMS

---

• https://arxiv.org/abs/2007.12099

---

[FilipLangr]

@AlexeyAB I believe there's a minor mistake in:

• Coord Conv – should be used only for conv-1x1 in backbone

PP-YOLO paper says (at page 4):
"In order to reduce the loss of efficiency as much as possible, we do not change convolutional layers in backbone, and only replace the 1x1 convolution layer in FPN and the first convolution layer in detection head with CoordConv."

[AlexeyAB]

Fixed it.

[toplinuxsir]

@AlexeyAB Great work, Where to find the pp yolo config file to train custom detection? Thanks !

[scamianbas]

@AlexeyAB Great work, Where to find the pp yolo config file to train custom detection? Thanks !

Hi @AlexeyAB and congrats for the great work. Is this still in the todo list ? Thanks !

[Goru1890]

Great work! @scamianbas Yes, it's seems to be still in progress https://github.com/AlexeyAB/darknet/projects/1#card-42538662.

[matankley]

Hi,
The new_coords=1 parameter in yolo_layer implements the calculation process in the [yolo_layer].

But from what I understand it doesn't implement CoordConv the same way as pp-yolo.
No extra 2 channels are added and there is no possibility to add new_coords toi [conv] layer.

Am i correct ?
Any plans to support this ?

Thanks

[tamir12312]

Hi,

I need help to understand how can I implement deformable convolution layer instead of convolution.
I can't find any parameter that suggests to replace the simple convolution operation with the deformable one.

In your comment, you wrote that "we replace the 3×3 convolutional layer in the last stage of ResNet with deformable convolution layer" - How did you do that and can I replicate it?

Thanks alot in advance!
Tamir :)