What is the reason for choosing [3,3,8,12] CGB layers for backbone?

[qaz670756]

The [3, 3, 8, 12] CGB layers was determined by experiments, which is common in other CNN designing process such as ResNet and CGNet. The principle is to use fewer layers in shallow levels with higher feature map resolution, and use more layers in deeper levels for extracting semantic information, which can reduce the network GFLOPS. You could try to varify them according to your task.

[heifanfanfan]

The [3, 3, 8, 12] CGB layers was determined by experiments, which is common in other CNN designing process such as ResNet and CGNet. The principle is to use fewer layers in shallow levels with higher feature map resolution, and use more layers in deeper levels for extracting semantic information, which can reduce the network GFLOPS. You could try to varify them according to your task.

Since I just came into contact with this research direction, can I add other contact information of the author to consult relevant questions