The Pytorch Implementation of Real-time Semantic Segmentation via Spatial-detail Guided Context Propagation

Dataset Set

Camvid

The Camvid dataset contains 701 images and 32 different semantic categories. We follow the previous works, and evaluate our segmentation model using 11 categories. For the details, please refer to the papers [1], [2], [3]. The Camvid dataset can be downloaded from http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/.

Cityscapes

The Cityscapes dataset contains 25 000 road scene images, in which 5000 images are finely annotated and 20 000 images are labelled with coarse annotation. In our experiments, we only adopt the fine-annotated subset. The fine-annotated subset involves 30 semantic categories. But we follow the previous works [2], [3], [4], [5], and adopt 19 categories in model evaluation. The Cityscapes dataset can be downloaded from http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/. Accordingly, the toolkits for data pre-processing can be found at https://github.com/mcordts/cityscapesScripts.

[1] Combining Appearance and Structure from Motion Features for Road Scene Understanding
[2] ICNet for real-time semantic segmentation on high-resolution images
[3] Efficient dense modules of asymmetric convolution for real-time semantic segmentation
[4] Pyramid scene parsing network
[5] PSANet: Point-wise spatial attention network for scene parsing

Main Environment

python==3.6.2
pytorch==1.1
numpy==1.15
torchvision==0.3.0
pillow==7.1.2
cython==0.29.20
scipy==1.1.0
scikit-learn==0.16.2

Script Running

(1). Before running the scripts, please run python src/setup.py build_ext --build-lib=./src/.

(2). For training the segmentation model, please run the command: python src/train.py --evaluate [False/True].

(3). We have saved the well-trained model on the Cityscapes dataset in Google Drive for reproducing our results. Specifically, firstly, put the download checkpoint into the folder ./ckpt. Secondly, run the command python src/test.py. The predictions for the test images are saved into the folder ./ckpt/test_result. Thirdly, please run the command zip -r test_result.zip test_result in bash. Finally, submit the file test_result.zip to the official online evaluator (https://www.cityscapes-dataset.com/submit/) to get the final performance in the Cityscapes' test set.

The results should be around the following:

Class level

Metric	Average	Road	Sidewalk	Building	Wall	Fence	Pole	Trafficilight	Trafficsign	Vegetation	Terrain	Sky	Person	Rider	Car	Truck	Bus	Train	Motorcycle	Bicycle
IoU (%)	70.9003	98.0907	82.6953	90.8128	46.2016	48.8514	56.3521	61.3421	68.3546	92.0528	69.1465	94.5957	79.5957	61.4092	93.6822	53.2762	69.9236	60.5087	53.1659	67.0485
iIoU (%)	43.4520	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	56.2162	35.7356	85.4904	22.7908	38.0693	29.0138	29.1411	51.1591

Category level

Metric	Average	Flat	Nature	Object	Sky	Construction	Human	Vehicle
IoU (%)	87.3518	98.3475	91.8220	63.0371	94.5957	91.0540	79.8839	92.7222
iIoU (%)	70.3811	n/a	n/a	n/a	n/a	n/a	57.425	83.371

If our work is helpful for your research, please consider citing our paper:

@article{hao2022real,
  title={Real-Time Semantic Segmentation via Spatial-Detail Guided Context Propagation},  
  author={Hao, Shijie and Zhou, Yuan and Guo, Yanrong and Hong, Richang and Cheng, Jun and Wang, Meng},  
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2022},  
  publisher={IEEE}  
}

Very thanks for the great works DeepLab, Condense, HyperSeg, Mobilenet, OpCounter, and TensorRT.

slothme / SGCPNet