trellixvulnteam / rch_sd76

RCH

code for our paper "Renovate Yourself: Calibrating Feature Representation of Misclassified Pixels for Semantic Segmentation"
the corresponding article

PipleLine of RCH

Results

Network	Encoder	Iteration	Train	Test	\rho	\tau	mIoU(log)	pth	comments
OCRNet	HRNet-W48	40000	train set	val set	0.9	0.5	82.24	pth
OCRNet	HRNet-W48	40000	train set	val set	0.9	0.5	82.29	pth	Reproduced
OCRNet	HRNet-W48	40000	train set	val set	0.8	0.5	82.25	_
OCRNet	HRNet-W48	40000	train set	val set	0.9	0.25	81.88	_
FCN	HRNet-W48	40000	train set	val set	0.9	0.25	81.83	pth

Requirements

python>=3.6
torch>=1.7
torchvision
visdom
numpy
pillow
scikit-learn
For the GPUs, we use 8 NVIDIA TITAN X(P) GPUs.

Usage

• core code of RCH is ./mmseg/models/decode_heads/decode_head_new.py
• Please replace your dataset dir into the corresponding config file.
• train OCRNet on cityscapes:

./tools/dist_train.sh ./configs/ocrnet/ocrnet_hr48_512x1024_40k_cityscapes.py 8

Now, you can use RCH to train OCRNet, FCN, DeepLab. We will continue to update more available decoders。

Citation

@InProceedings{wang2021rch,
tilte = {Renovate Yourself: Calibrating Feature Representation of Misclassified Pixels for Semantic Segmentation},
author = {Wang, Hualiang and Chu, Huanpeng and Fu, siming},
booktitle = {AAAI},
year = {2022}
}