Are the step settings in `ssd_300_mobilenet0_25_coco` in GluonCV model_zoo correct?

This code line is here:
ssd_300_mobilenet0.25_coco

Firstly, the first scale feature map is drawn from relu22_fwd from mobilenetv1 , however when input data is 300 * 300, this layer’s feature map is 19 * 19, or actually 18 * 18 using mobilenet.py in model_zoo folder, which means that the step is actually 16.

Then, the step parameter in function ssd_300_mobilenet0.25_coco is started from 8, and remains are [8, 16, 32, 64, 100, 300] . I noted that the steps are used within anchor generation, and I want to know the step settings here could impact the training accuracy (mAP)?

thank you.

@Sergey

Please have a look? thankyou very much.

Hey @TriLoon, I think @thomelane can better help you with that.

Hi @TriLoon,

Could you confirm your exact issue here. Are you saying that your feature maps get reduced by a factor of 16 in the first stage, but the anchor boxes only reduce by a factor of 8?

Yes, the truth is that the first stage feature map get reduced by a factor of 16 in SSD from gluoncv/model_zoo/ssd/presets.py.

Model name are: ssd_mobilenet0.25_300_coco and others.