RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024)

1. Install Environment, Datasets, Pretrained Model

1.1 Environment

Our project is based-on MMSegmentation. Please follow the official MMsegmentation INSTALL.md with Python= 3.7, Pytorch= 1.12.1, CUDA=11.3 and CUDNN=8.3.2.

1.2 Datasets

NYUD:

Download the augmented NYUD data from EDTER Repository:

-Download Train Data

-Download Test Data

Put these files into data/NYUD/.

BSDS:

Download the augmented BSDS data from here

Put these file into data/BSDS_RS/.

Extract these files via tar zxvf filename.tar.gz

1.3 Initial Weights

Download the pretrained model from here

Put this file into preTrain/.

2. Train Model

2.1 NYU:

-Change delta as 0.4 and split (in Line 9) based on your GPU memory (split=1 requires huge memory about ~40 GB) in mmseg/model/losses/ap_loss.py

-Run the following command to start training.

python tools/train.py configs/APLoss/base_320_fullData.py --options model.pretrained=preTrain/swin_base_patch4_window12_384_22k.pth model.backbone.use_checkpoint=True --work-dir your_folder

2.2 BSDS:

For Only Ranking:

-Change delta as 0.1 and split (in Line 9) based on your GPU memory (split=1 requires huge memory about ~40 GB) in mmseg/model/losses/ap_loss.py

-Run the following command to start training.

python tools/train.py configs/APLoss/base_320_fullData_bsds.py --options model.pretrained=preTrain/swin_base_patch4_window12_384_22k.pth model.backbone.use_checkpoint=True --work-dir your_folder

For Ranking & Sorting:

-Change delta as 0.1 and split (in Line 9) based on your GPU memory (split=1 requires huge memory about ~40 GB) in mmseg/model/losses/rank_loss.py and mmseg/model/losses/sort_loss.py

• Comment the line 12 in mmseg/datasets/pipelines/bsds_get_gtfiles.py

-Run the following command to start training.

python tools/train.py configs/RSLoss/base_320_fullData_bsds.py --options model.pretrained=preTrain/swin_base_patch4_window12_384_22k.pth model.backbone.use_checkpoint=True --work-dir your_folder

3. Inference

-Run the following command to start inference. python tools/test.py --config configs/APLoss/base_320_fullData_bsds.py --checkpoint your_folder/xxx.pth --tmpdir your_save_result_dir

4. Pre-trained Models:

-NYUD (Only Ranking)

-BSDS (Ranking & Sorting)

4. Acknowledgements

Thanks to the previous open-sourced repo:

EDTER

Swin-Tranformer

MMSegmentation

5. Reference

@InProceedings{cetinkaya2024ranked,
  title={RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses}, 
  author={Bedrettin Cetinkaya and Sinan Kalkan and Emre Akbas},
  year={2024},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  url={https://ranked-cvpr24.github.io/}
}