Visual Attention Network (VAN) for Segmentaion

This repo is a PyTorch implementation of applying VAN (Visual Attention Network) to semantic segmentation. The code is based on mmsegmentaion.

More details can be found in Visual Attention Network.

Citation

@article{guo2022visual,
  title={Visual Attention Network},
  author={Guo, Meng-Hao and Lu, Cheng-Ze and Liu, Zheng-Ning and Cheng, Ming-Ming and Hu, Shi-Min},
  journal={arXiv preprint arXiv:2202.09741},
  year={2022}
}

Results

Notes: Pre-trained models can be found in TsingHua Cloud.

VAN + UperNet

Method	Backbone	Pretrained	Iters	mIoU(ms)	Params	FLOPs	Config	Download
UperNet	VAN-B0	IN-1K	160K	41.1	32M	-	config	-
UperNet	VAN-B1	IN-1K	160K	44.9	44M	-	config	-
UperNet	VAN-B2	IN-1K	160K	50.1	57M	948G	config	TsingHua Cloud
UperNet	VAN-B3	IN-1K	160K	50.6	75M	1030G	config	TsingHua Cloud
UperNet	VAN-B4	IN-1K	160K	52.2	90M	1098G	config	TsingHua Cloud
UperNet	VAN-B4	IN-22K	160K	53.5	90M	1098G	config	TsingHua Cloud
UperNet	VAN-B5	IN-22K	160K	53.9	117M	1208G	config	TsingHua Cloud
UperNet	VAN-B6	IN-22K	160K	54.7	231M	1658G	config	TsingHua Cloud

Notes: In this scheme, we use multi-scale validation following Swin-Transformer. FLOPs are tested under the input size of 2048 $\times$ 512 using torchprofile (recommended, highly accurate and automatic MACs/FLOPs statistics).

VAN + Semantic FPN

Backbone	Iters	mIoU	Config	Download
VAN-Tiny	40K	38.5	config	Google Drive
VAN-Small	40K	42.9	config	Google Drive
VAN-Base	40K	46.7	config	Google Drive
VAN-Large	40K	48.1	config	Google Drive

Preparation

Install MMSegmentation and download ADE20K according to the guidelines in MMSegmentation.

Requirement

pip install mmsegmentation==0.26.0 (https://github.com/open-mmlab/mmsegmentation/tree/v0.26.0)

Training

We use 8 GPUs for training by default. Run:

./dist_train.sh /path/to/config 8

Evaluation

To evaluate the model, run:

./dist_test.sh /path/to/config /path/to/checkpoint_file 8 --eval mIoU

FLOPs

Install torchprofile using

pip install torchprofile

To calculate FLOPs for a model, run:

bash tools/flops.sh /path/to/checkpoint_file --shape 512 512

Acknowledgment

Our implementation is mainly based on mmsegmentaion, Swin-Transformer, PoolFormer, and Enjoy-Hamburger. Thanks for their authors.

LICENSE

This repo is under the Apache-2.0 license. For commercial use, please contact the authors.