Skip to content

Latest commit

 

History

History
31 lines (25 loc) · 1.08 KB

README.md

File metadata and controls

31 lines (25 loc) · 1.08 KB

PointRend-TF

PointRend dnn layers implement in Tensorflow 1.15

  • This Head layers should put after the mask head of mask-rcnn (crop_bbox - >roi_Align -> mask_logits).
  • The mask size 7x7 is default mask head predict, which were 14x14 downsapmling from 2x2 conv2d.
  • This project is default predict 2 classes cls1/background0.
  • Only take features from last fpn layer.

example

from PointRendHead import PointHead_Branch
### put this layer after your mask-rcnn fpn_features -> rpn -> proposal -> roi_align layers
pointhead = PointHead_Branch(pred_classes, crop_rois, msk_logit, fpn_feats, is_training)

### training
#pred_cls (from mask-rcnn)
#gt_masks (from dataset)
msk_logit = pointhead.point_msk_logits
point_coords = pointhead.point_coords
### loss
loss = pointhead.roi_mask_point_loss(pred_cls, msk_logit, points_coord, gt_masks)

### evaluation
pointhead.point_rend_mask_logit

The whole algo are reference to:
detectron2.
PointRend: Image Segmentation as Rendering.