rcf

variety of Richer Convolutional Features for Edge Detection (resnet101-based)

results

I test two types of loss: last layer loss / sum up each layer's loss. Detail here.

1. last layer only: ODS: 0.8135 , OIS: 0.831 on BSDS500 dataset pretrained model, last layer only

2. all layers: ODS: 0.796 , OIS: 0.814 on BSDS500 dataset pretrained model, all layers

requirements

• pytorch 0.4.1
• python 3.6.6
• dataset(provide by original repo)
  • http://mftp.mmcheng.net/liuyun/rcf/data/bsds_pascal_train_pair.lst
  • http://mftp.mmcheng.net/liuyun/rcf/data/HED-BSDS.tar.gz
  • http://mftp.mmcheng.net/liuyun/rcf/data/PASCAL.tar.gz
• and other requirements... (cv2, numpy , etc.)

usage

train:

• put your data in 'data/HED-BSDS_PASCAL' (or make a soft link)
• python train.py

test:

• python test.py

a simple example:

• python demo.py

evaluate:

it may take several hours...

• pretrained model, last layer only or pretrained model, all layers
• requirements:
  • matlab
  • hed link
• you should modify path to your predicts and path to ground truth (.mat)
  • path to your predicts should contain two folders: png and mat
• sh eval.sh

related projects

• rcf pytorch implementation
• rcf
• hed pytorch implementation