Reparameterizable Dual-Resolution Network for Real-time Semantic Segmentation (arXiv)
By Guoyu Yang, Yuan Wang, Daming Shi*
This project is based on mmsegmentation.
python==3.8.10
pytorch==1.12.1
torchvision==0.13.1
mmengine==0.7.3
mmcv==2.0.0
mmsegmentation==1.0.0
Please refer to mmsegmentation for installation.
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.
Single gpu for train:
CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/train.py rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py --work-dir ./weight/segMultiple 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/segTrain in pycharm: If you want to train in pycharm, you can run it in train.py.
see more details at mmsegmentation.
CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/test.py rdrnet-s-simple_2xb6-120k_cityscapes-1024x1024.py ./weight/seg/rdrnet_weight.pthTest in pycharm: If you want to test in pycharm, you can run it in test.py.
see more details at mmsegmentation.
| 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.
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}
}