🚀 If it helps you, click a star! ⭐

Update log

• 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 .....

1. Display (Cityscapes)

• 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

2. Install

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
  

3. Model

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.......

4. Data preprocessing

This project enables you to expose data sets: Cityscapes、ISPRS
The data set is uploaded later .....
Cityscapes data set preparation is shown here:

4.1 Download the dataset

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

4.2 Onehot encoding of label image

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
                        |--*******

4.3 Generate image path

• Generate a txt containing the image path
  Use script dataset/generate_txt.py to generate the path txt file containing the original image and labels. A total of 3 txt files 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 txt are 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 txt are shown as follows:

origin image path + the separator '\t' + label path +  the separator '\n'

TODO.....

5. How to train

sh train.sh

5.1 Parameters

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

6. How to validate

sh predict.sh