How should I adjust the training parameters, such as base_anchor_shapes?

[fanhenghui]

I am currently researching the detection sub-project. When training for lung nodules smaller than 3mm, how should I adjust the training parameters, such as base_anchor_shapes?

[KumoLiu]

Hi @fanhenghui, adjust base_anchor_shapes to include smaller anchor box sizes that correspond to the pixel scale of 3mm nodules in your images. If your image resolution makes 3mm correspond to, say, 5-7 pixels, set anchors to sizes close to these dimensions.

tutorials/detection/luna16_training.py

Lines 164 to 165 in fc15cdd

	# base_anchor_shapes: anchor shape for the most high-resolution output,
	# when target boxes are small, set it smaller