Incorrect yolo head anchor box mask ? - Githubissues

Tianxiaomo / pytorch-YOLOv4

PyTorch ,ONNX and TensorRT implementation of YOLOv4

Apache License 2.0

4.48k stars 1.49k forks source link

Incorrect yolo head anchor box mask ? #264

Open pop134 opened 4 years ago

pop134 commented 4 years ago

As in the models.py:

https://github.com/Tianxiaomo/pytorch-YOLOv4/blob/4ccef0ec8fe984e059378813e33b3740929e0c19/models.py#L446-L448

The x20 is the biggest feature map output after the whole FPN block. Then it's fed to yolo head which take the name input1.

https://github.com/Tianxiaomo/pytorch-YOLOv4/blob/4ccef0ec8fe984e059378813e33b3740929e0c19/models.py#L370

After that, the input1 (x20 feature map) after go through 2 conv layers is matched with yolo1 - which takes the anchor masks = [0,1,2] (smallest anchors). https://github.com/Tianxiaomo/pytorch-YOLOv4/blob/4ccef0ec8fe984e059378813e33b3740929e0c19/models.py#L331-L334

I'm I understanding this right? Cause I think the anchors masks here must be [6,7,8] due to the size of the features to detect small objects. Big thanks.

matankley commented 4 years ago

Hi @pop134, This confused me as well, Did you manage to get better results by modifying yolo1 to hold the biggest stride and anchor mask ? Thanks