Caffe: why prior_box_layer has no backpropagation implementation?

Created on 6 Jul 2016  路  2Comments  路  Source: weiliu89/caffe

I don't quite understand how prior_box_layer performs backpropagation if it is not implemented

picture mmmonkeytao

Most helpful comment

prior_box_layer is only used to generate prior (default) bbox, which is fixed. The location prediction (e.g. *_mbox_loc) is the real prediction (offset on top of prior box), which is backpropagated during training.

picture weiliu89 on 6 Jul 2016
👍2

All 2 comments

prior_box_layer is only used to generate prior (default) bbox, which is fixed. The location prediction (e.g. *_mbox_loc) is the real prediction (offset on top of prior box), which is backpropagated during training.

weiliu89 picture weiliu89 on 6 Jul 2016
👍2

I see. In DetectionOutput layer, the fixed prior bbox is required in the loss caculation between predicted bbox and ground truth bbox.

mmmonkeytao picture mmmonkeytao on 6 Jul 2016
👍1
Was this page helpful?
0 / 5 - 0 ratings

Related issues

lcj1105 picture lcj1105  路  3Comments
dakatapetrov picture dakatapetrov  路  5Comments
Jiaqiww picture Jiaqiww  路  5Comments
ShadowLau picture ShadowLau  路  4Comments
Coderx7 picture Coderx7  路  3Comments