- 2020.12.10 Project structure adjustment, the previous code has been deleted, the adjustment will be re-uploaded code
- 2021.04.09 Re-upload the code, "V1 Commit"
- 2021.04.22 update torch distributed training
- Ongoing update .....
- Using model DDRNet 1525 test sets, official MIOU =78.4069%
 Average results |
 Class results1 |
 Class results2 |
 Class results3 |
- Comparison of the original and predicted images
 origin |
 label |
 predict |
pip install -r requirements.txt
Experimental environment:
- Ubuntu 16.04 Nvidia-Cards >= 1
- python==3.6.5
- See Dependency Installation Package for details in requirement.txt
All the modeling is done in builders/model_builder.py
- FCN
- FCN_ResNet
- SegNet
- UNet
- BiSeNet
- BiSeNetV2
- PSPNet
- DeepLabv3_plus
- HRNet
- DDRNet
| Model | Backbone | Val mIoU | Test mIoU | Imagenet Pretrain | Pretrained Model |
|---|---|---|---|---|---|
| PSPNet | ResNet 50 | 76.54% | - | √ | PSPNet |
| DeeplabV3+ | ResNet 50 | 77.78% | - | √ | DeeplabV3+ |
| DDRNet23_slim | - | DDRNet23_slim_imagenet | |||
| DDRNet23 | - | DDRNet23_imagenet | |||
| DDRNet39 | - | 79.63% | - | DDRNet39_imagenet | DDRNet39 |
| Updating more model....... |
This project enables you to expose data sets: Cityscapes、ISPRS
The data set is uploaded later .....
Cityscapes data set preparation is shown here:
Download the dataset from the link on the website, You can get *leftImg8bit.png suffix of original image under folder leftImg8bit,
a) *color.png、b) *labelIds.png、c) *instanceIds.png suffix of fine labeled image under folder gtFine.
*leftImg8bit.png : the origin picture
a) *color.png : the class is encoded by its color
b) *labelIds.png : the class is encoded by its ID
c) *instanceIds.png : the class and the instance are encoded by an instance ID
The real label gray scale image Onehot encoding used by the semantic segmentation task is 0-18, so the label needs to be encoded.
Using scripts dataset/cityscapes/cityscapes_scripts/process_cityscapes.py
to process the image and get the result *labelTrainIds.png.
process_cityscapes.py usage: Modify 486 lines `Cityscapes_path'is the path to store your own data.
- Comparison of original image, color label image and gray label image (0-18)
 ***_leftImg8bit |
 ***_gtFine_color |
 ***_gtFine_labelTrainIds |
- Local storage path display
/data/open_data/cityscapes/:
data
|--open_data
|--cityscapes
|--leftImg8bit
|--train
|--cologne
|--*******
|--val
|--*******
|--test
|--*******
|--gtFine
|--train
|--cologne
|--*******
|--val
|--*******
|--test
|--*******
- Generate a txt containing the image path
Use scriptdataset/generate_txt.pyto generate the pathtxtfile containing the original image and labels. A total of 3txtfiles will be generated:cityscapes_train_list.txt、cityscapes_val_list.txt、cityscapes_test_list.txt, and copy the three files to the dataset root directory.
data
|--open_data
|--cityscapes
|--cityscapes_train_list.txt
|--cityscapes_val_list.txt
|--cityscapes_test_list.txt
|--leftImg8bit
|--train
|--cologne
|--*******
|--val
|--*******
|--test
|--*******
|--gtFine
|--train
|--cologne
|--*******
|--val
|--*******
|--test
|--*******
- The contents of the
txtare shown as follows:
leftImg8bit/train/cologne/cologne_000000_000019_leftImg8bit.png gtFine/train/cologne/cologne_000000_000019_gtFine_labelTrainIds.png
leftImg8bit/train/cologne/cologne_000001_000019_leftImg8bit.png gtFine/train/cologne/cologne_000001_000019_gtFine_labelTrainIds.png
..............
- The format of the
txtare shown as follows:
origin image path + the separator '\t' + label path + the separator '\n'
sh train.sh
python -m torch.distributed.launch --nproc_per_node=2 \
train.py --model PSPNet_res50 --out_stride 8 \
--max_epochs 200 --val_epochs 20 --batch_size 4 --lr 0.01 --optim sgd --loss ProbOhemCrossEntropy2d \
--base_size 768 --crop_size 768 --tile_hw_size 768,768 \
--root '/data/open_data' --dataset cityscapes --gpus_id 1,2
sh predict.sh
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent