CSPDarknet53 as a Backbone

[ErikValle]

The YOLOv4 paper suggests using a backbone based on CSPResNext50 or CSPDarknet53. The first one has a higher accuracy when trained with different features, while the other shows higher accuracy in object detection. Since the yolov5x YAML file follows a kind of CSPDarknet53, I expected to see something following this architecture but with CSP. Nevertheless, it has the elements below:

YOLOv5 backbone

backbone:
[[-1, 1, Focus, [64, 3]], # 0-P1/2
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4
[-1, 3, C3, [128]],
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8
[-1, 9, C3, [256]],
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16
[-1, 9, C3, [512]],
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
[-1, 1, SPP, [1024, [5, 9, 13]]],
[-1, 3, C3, [1024, False]], # 9
]

Questions:

1. Shouldn't the backbone include a 32-depth Conv block as listed below? I tried to adjust the architecture as in CSPDarknet53 but it may be wrong

YOLOv5 backbone (my suggestion)

backbone:
[[-1, 1, Conv, [32, 3]], # 0-P1/2
[-1, 1, Conv, [64, 3, 2]], # 1-P2/4
[-1, 3, C3, [128]],
[-1, 1, Conv, [128, 3, 2]], # 3-P3/8
[-1, 9, C3, [256]],
[-1, 1, Conv, [256, 3, 2]], # 5-P4/16
[-1, 9, C3, [512]],
[-1, 1, Conv, [512, 3, 2]], # 7-P5/32
[-1, 9, C3, [1024]],
[-1, 1, Conv, [1024, 3, 2]], # 9-P6/64
[-1, 9, C3, [2048]],
[-1, 1, SPP, [2048, [5, 9, 13]]],
[-1, 3, C3, [1024, False]], # 11
]
2. Why did you use Focus as the first operation instead of Conv? I was just curious, you guys have already discussed it here: #4825
Thank you

[glenn-jocher]

@ErikValle YOLOv5 backbone and head have evolved over releases and include C3 modules which are reduced CSP modules.

Focus conversation is fully covered in #4825

[ErikValle]

Thank you!