Darknet: YOLACT: Add real-time Instance Segmentation

[AlexeyAB]

Add real-time Instance Segmentation:

• Look at: https://github.com/pjreddie/darknet/blob/master/examples/instance-segmenter.c

• YOLACT++ https://arxiv.org/abs/1912.06218v1

• YOLACT (You Only Look At CoefficienTs): https://arxiv.org/abs/1904.02689v1 and https://github.com/dbolya/yolact

  1. Add additional output-prototype-layer with full-resolution masks (W x H x K) with upsampling_layers=subsampling_layers, where is K=32 number of prototype masks

  2. Use (4+1+classes+K)*anchors outputs of Detector instead of (4+1+classes)*anchors. Then we do NMS or Fast-NMS. Then we do GEMM between output-prototype-layer [W x H][K] and found N-bboxes [N][K] to get [WxH][N] masks for N found bboxes, and apply logistic_activation.

During training: to compute mask loss, we simply take the pixel-wise binary cross entropy between assembled masks M and the ground truth masks. We crop with the ground truth bounding box and divide Lmask by the ground truth bounding box area.

Test on:

• MS COCO
• Kitty
• Cityscapes

[LukeAI]

This is (another) exciting and impressive potential enhancement.
I was wondering, @AlexeyAB - how do you find time to do all this stuff and then give it away for free - are you being paid to produce this stuff or are you a passionate hobbyist?

[abhigoku10]

@AlexeyAB have you implemented the yolAct architecture in the instance segmenter.c ??

[zpmmehrdad]

Hi @AlexeyAB,

I really need it. When do you think we can use this feature?

Thanks in advance.

[LukeAI]

Why not use https://github.com/dbolya/yolact ?

[abhigoku10]

@zpmmehrdad suggest to use @LukeAI has suggest its good in fps and is implemented in pytorch

[zpmmehrdad]

@abhigoku10 Hi,

Thanks for suggesting, I want to use it in C++.

[abhigoku10]

@zpmmehrdad yes even i wanted to use c++ later i shifted my whole development environment pls let me know if come across anyother repo

[WongKinYiu]

CenterMask:Real-Time Anchor-Free Instance Segmentation: https://arxiv.org/pdf/1911.06667.pdf

Faster and better than YOLACT.

[Kyuuki93]

@AlexeyAB

YOLACT++: Better Real-time Instance Segmentation

• paper https://arxiv.org/abs/1912.06218
• code https://github.com/dbolya/yolact (same with yolact)

| Method| mAP | FPS |
| --- | --- | ---|
|yoloact | 29.8%| 33.0|
|yoloact++|34.1%| 33.5|

The different from yolact:

1. Add a fast mask re-scoring branch +1.1%
2. Deformable convolution +1.5%
3. Search a better aspect ratio for each FPN layer
4. More anchor