This project provides an implementation for "Learning Dynamic Routing for Semantic Segmentation" (CVPR2020 Oral) on PyTorch. For the reason that experiments in the paper were conducted using internal framework, this project reimplements them on dl_lib and reports detailed comparisons below. Some parts of code in dl_lib are based on detectron2.
- Python >= 3.6
python3 --version
- PyTorch >= 1.3
pip3 install torch torchvision
- OpenCV
pip3 install opencv-python
- GCC >= 4.9
gcc --version
Make sure that your get at least one gpu when compiling. Run:
git clone https://github.com/yanwei-li/DynamicRouting.gitcd DynamicRoutingsudo python3 setup.py build develop
We use Cityscapes dataset for training and validation. Please refer to datasets/README.md or dataset structure in detectron2 for more details.
- Cityscapes Download
We give ImageNet pretained models:
- Layer16-Fix GoogleDrive
- Layer33-Fix GoogleDrive
For example, if you want to train Dynamic Network with Layer16 backbone:
- Train from scratch
cd playground/Dynamic/Seg.Layer16 dl_train --num-gpus 4 - Use ImageNet pretrain
cd playground/Dynamic/Seg.Layer16.ImageNet dl_train --num-gpus 4 MODEL.WEIGHTS /path/to/your/save_dir/ckpt.pth
NOTE: Please set FIX_SIZE_FOR_FLOPS to [768,768] and [1024,2048] for training and evaluation, respectively.
You can evaluate the trained or downloaded model:
- Evaluate the trained model
dl_test --num-gpus 8
- Evaluate the downloaded model:
dl_test --num-gpus 8 MODEL.WEIGHTS /path/to/your/save_dir/ckpt.pth
NOTE: If your machine does not support such setting, please change settings in config.py to a suitable value.
Without ImageNet Pretrain:
| Methods | Backbone | Iter/K | mIoU (paper) | GFLOPs (paper) | mIoU (ours) | GFLOPs (ours) | Model |
|---|---|---|---|---|---|---|---|
| Dynamic-A | Layer16 | 186 | 72.8 | 44.9 | 73.9 | 52.5 | GoogleDrive |
| Dynamic-B | Layer16 | 186 | 73.8 | 58.7 | 74.3 | 58.9 | GoogleDrive |
| Dynamic-C | Layer16 | 186 | 74.6 | 66.6 | 74.8 | 59.8 | GoogleDrive |
| Dynamic-Raw | Layer16 | 186 | 76.1 | 119.5 | 76.7 | 114.9 | GoogleDrive |
| Dynamic-Raw | Layer16 | 558 | 78.3 | 113.3 | 78.1 | 114.2 | GoogleDrive |
With ImageNet Pretrain:
| Methods | Backbone | Iter/K | mIoU (paper) | GFLOPs (paper) | mIoU (ours) | GFLOPs (ours) | Model |
|---|---|---|---|---|---|---|---|
| Dynamic-Raw | Layer16 | 186 | 78.6 | 119.4 | 78.8 | 117.8 | GoogleDrive |
| Dynamic-Raw | Layer33 | 186 | 79.2 | 242.3 | 79.4 | 243.1 | GoogleDrive |
- Faster inference speed
- Support more vision tasks
- Object detection
- Instance segmentation
- Panoptic segmentation
Consider cite the Dynamic Routing in your publications if it helps your research.
@inproceedings{li2020learning,
title = {Learning Dynamic Routing for Semantic Segmentation},
author = {Yanwei Li, Lin Song, Yukang Chen, Zeming Li, Xiangyu Zhang, Xingang Wang, Jian Sun},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
year = {2020}
}
Consider cite this project in your publications if it helps your research.
@misc{DynamicRouting,
author = {Yanwei Li},
title = {DynamicRouting},
howpublished = {\url{https://github.com/yanwei-li/DynamicRouting}},
year ={2020}
}