CBL-Conditional-Boundary-Loss

The official implementation of the accepted IEEE-TIP paper Conditional Boundary Loss for Semantic Segmentation.
We provide the code of our CBL based on our_maskformer and MMsegmentation. The whole CBL MMsegmentation code base will be uploaded soon. For now, your can use our CBL in MMsegmentation following the instructions down below.

Our Results

More results of other models like Swin-B and PSP, together with the trained model weights file, will be updated soon once we finished the orgnization of our project.
Temporary result table:

Cityscapes

model	Backbone	iter	Crop Size	mIoU MMseg (single scale)	mIoU +CBL Ours (single scale)
OCRNet	HRNetW48	80K	512x1024	80.70	81.95

ADE20K

model	Backbone	mIoU(SS)	mIoU(MS)
MaskFormer	Swin-B	--	53.83(official)
MaskFormer +CBL	Swin-B	53.49	54.89 Trained Model Code:CBL0
Mask2Former	Swin-B	--	55.07(official)
Mask2Former +CBL	Swin-B	54.79	56.05 Trained Model Code:CBL0

How to train MaskFormer +CBL model

We build our implementation based on the official code base of MaskFormer. Please refer to our MaskFormer code base at our_maskformer This implementation enables easy reproduction of our CBL on MaskFormer, which do not need the above complicated steps for mmsegmentation. The trained MaskFormer+CBL model can also be found at MaskFormer+CBL Trained Model Code:CBL0 The trained Mask2Foremer+CBL model is also provided at Mask2Former+CBL Trained Model Code:CBL0

How to use our code in MMsegmentation

We follow the implementation of MMsegmentation. Here we provide the code of CBL based on the OCRHead in CBLocr_head.py.
The class name of the OCRHead with our CBL is New_ER5OCRHead.

Download our code and add the CBLocr_head.py to Path mmseg/models/decode_head/ocr_head.py in your MMsegmentation source code project.
Import the ER5OCRHead class in the mmseg/models/decode_head/init.py:
change the line from .ocr_head import OCRHead as from .ocr_head import OCRHead, New_ER5OCRHead
Add the code for generating Ground Truth boundary for training:
1. Download the biou.py and add it to mmseg/core/evaluation/biou.py
2. Import the functions in biou.py:
  add the following line to the mmseg/core/evaluation/init.py
  from .biou import multi_class_gt_to_boundary
  then add 'multi_class_gt_to_boundary'to the list of __all__ = [xxx]
3. Download the boundary.py and add it to mmseg/datasets/pipelines/boundary.py
4. Import the GenerateBoundary class in the mmseg/datasets/pipelines/init.py:
  add the following line to the mmseg/datasets/pipelines/init.py
  from .boundary import GenerateBoundary
  then add 'GenerateBoundary' to the list of __all__ = [xxx]
Using our config.py to train a OCRNet.
For example, to train a OCRNet-HRNetW48 on cityscapes, please run the following code:
sh tools/dist_train.sh YOUR_PATH_TO_THE_CONFIG/erocrnet_hr48_512x1024_80k_cityscapes_fp16.py 8

TO-DO List(after accepted)

1.Upload the whole CBL project based on MMsegmentation (including CBL trained models with PSPNet, DeeplabV3+, Swin-B)

2.~Upload the whole CBL project based on MaskFormer (including CBL trained models with MaskFormer)~ (Done)

3.Write a new instruction about how to run our CBL on the above-mentioned projects.

dywu98 / CBL-Conditional-Boundary-Loss