Reparameterizable Dual-Resolution Network for Real-time Semantic Segmentation (arXiv)

By Guoyu Yang, Yuan Wang, Daming Shi*

This project is based on mmsegmentation.

Highlight

The trade-off between inference speed and accuracy for real-time semantic segmentation models on the Cityscapes test set.

Experiment

Environment

python==3.8.10
pytorch==1.12.1
torchvision==0.13.1
mmengine==0.7.3
mmcv==2.0.0
mmsegmentation==1.0.0

Install

Please refer to mmsegmentation for installation.

Dataset

RDRNet
├── mmsegmentation
├── figures
├── data
│   ├── cityscapes
│   │   ├── leftImg8bit
│   │   │   ├── train
│   │   │   ├── val
│   │   ├── gtFine
│   │   │   ├── train
│   │   │   ├── val
│   ├── CamVid
│   │   ├── train
│   │   ├── train_labels
│   │   ├── test
│   │   ├── test_labels
│   ├── VOCdevkit
│   │   ├── VOC2012
│   │   │   ├── JPEGImages
│   │   │   ├── SegmentationClass
│   │   │   ├── ImageSets
│   │   │   │   ├── Segmentation
├── rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py
├── train.py
├── test.py

Cityscapes could be downloaded from here. Camvid could be downloaded from here. Pascal VOC 2012 could be downloaded from here.

Training

Single gpu for train:

CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/train.py rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py --work-dir ./weight/seg

Multiple gpus for train:

CUDA_VISIBLE_DEVICES=0,1 bash ./mmsegmentation/tools/dist_train.sh rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py 2 --work-dir ./weight/seg

Train in pycharm: If you want to train in pycharm, you can run it in train.py.

see more details at mmsegmentation.

Testing

CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/test.py rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py ./weight/seg/rdrnet_weight.pth

Test in pycharm: If you want to test in pycharm, you can run it in test.py.

see more details at mmsegmentation.

Results on Cityscapes

Method	FPS	Params (M)	GFLOPs	ImageNet	val	test
BiSeNetV1	65.9	13.3	118.0	✓	74.4	73.6
BiSeNetV2	74.4	3.4	98.4	✗	73.6	72.2
DDRNet-23-Slim	131.7	5.7	36.3	✗	76.3	74.6
DDRNet-23	54.6	20.3	143.0	✗	78.0	77.6
PIDNet-S	102.6	7.7	47.3	✗	76.4	76.2
PIDNet-M	42.0	28.7	177.0	✗	78.2	78.3
PIDNet-L	31.8	37.3	275.0	✗	78.8	78.4
RDRNet-S-Simple	134.6	7.2	41.0	✗	76.7	75.4
RDRNet-S	129.4	7.3	43.4	✗	76.8	76.0
RDRNet-M	52.5	26.0	162.0	✗	78.8	78.3
RDRNet-L	39.0	36.9	238.0	✗	79.6	78.6

During the evaluation on the validation set, the training set was utilized for training. During the evaluation on the test set, both the training and validation sets were employed for training. When performing inference on RTX 3090, the image resolution used was 1024 ✕ 2048.

Citations

If you find RDRNet useful in your research, please consider citing:

@misc{yang2024reparameterizable,
      title={Reparameterizable Dual-Resolution Network for Real-time Semantic Segmentation}, 
      author={Guoyu Yang and Yuan Wang and Daming Shi},
      year={2024},
      eprint={2406.12496},
      archivePrefix={arXiv}
}