Skip to content


Proposal for Faster R-CNN.

This layer performs proposal roi presentation operation for spatial data.
It takes two inputs; cls_prob and bbox_pred.
The shape of cls_prob is (h, w, 2n_anchors).
The shape of bbox_pred is (h, w, 4
n_anchors); i.e. when the last dimension is 36, it has 4 x 9 (anchors).
The shape of output is (num_roi, 5) where num_roi is fixed value 300.


  • feat_stride : stride of feature map in input image coordinate (determined by model structure).
  • scales : list of scales of proposals
  • origin : String, original channel-position when importing from other deep-learnig frameworks.
    'chl' for channel last or 'chf' for channel first. The default is 'chl'.