DeepSegmentor
A Pytorch implementation of DeepCrack and RoadNet projects.
1.Datasets
Please download the corresponding dataset and prepare it by following the guidance.
2.Installation
We provide an user-friendly configuring method via Conda system, and you can create a new Conda environment using the command:
conda env create -f environment.yml
3.Balancing Weights
We follow the Median Frequency Balancing method, using the command:
python3 ./tools/calculate_weights.py --data_path <path_to_segmentation>
4.Training
Before the training, please download the dataset and copy it into the folder datasets.
- Crack Detection
sh ./scripts/train_deepcrack.sh <gpu_id>
- Road Detection
sh ./scripts/train_roadnet.sh <gpu_id>
We provide our pretrained models here:
| Model | Google Drive | Baidu Yun | Others |
|---|---|---|---|
| DeepCrack | Fine-tuned | ||
| RoadNet | Roughly trained | ||
| RoadNet++ | [link] | [link] | - |
5.Testing
- Crack Detection
sh ./scripts/test_deepcrack.sh <gpu_id>
| Image | Ground Truth | GF | fused | side1 | side2 | side3 | side4 | side5 |
|---|---|---|---|---|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
- Road Detection
sh ./scripts/test_roadnet.sh <gpu_id>
| Image | Ground Truth | Prediction |
|---|---|---|
 |
 |
 |
6.Evaluation
- Metrics (appeared in our papers):
| Metric | Description | Usage |
|---|---|---|
| P | Precision, TP/(TP+FP) |
segmentation |
| R | Recall, TP/(TP+FN) |
segmentation |
| F | F-score, 2PR/(P+R) |
segmentation |
| TPR | True Positive Rate, TP/(TP+FN) |
segmentation |
| FPR | False Positive Rate, FP/(FP+TN) |
segmentation |
| AUC | The Area Under the ROC Curve | segmentation |
| G | Global accuracy, measures the percentage of the pixels correctly predicted | segmentation |
| C | Class average accuracy, means the predictive accuracy over all classes | segmentation |
| I/U | Mean intersection over union | segmentation |
| ODS | the best F-measure on the dataset for a fixed scale | edge,centerline |
| OIS | the aggregate F-measure on the dataset for the best scale in each image | edge,centerline |
| AP | the average precision on the full recall range | edge,centerline |
Note: If you want to apply the standard non-maximum suppression (NMS) for edge/centerline thinning. Please see more details in Piotr's Structured Forest matlab toolbox or some helper functions provided in the hed/eval.
[See more details (Evaluation + Guided Filter + CRF) >>>]
Usage:
cd eval
python eval.py --metric_mode prf --model_name deepcrack --output deepcrack.prf
[Display the accuracy curves >>>]
Acknowledgment
-
This code is based on the pytorch-CycleGAN-and-pix2pix. Thanks to the contributors of this project.
-
If you are familar to Google Colab, there is an implementation of Colab version (provided by DZDL/DeepSegmentor). Besides, there is a WebApp version of this project: crack-detector.
References
If you take use of our datasets or code, please cite our papers:
@article{liu2019deepcrack,
title={DeepCrack: A Deep Hierarchical Feature Learning Architecture for Crack Segmentation},
author={Liu, Yahui and Yao, Jian and Lu, Xiaohu and Xie, Renping and Li, Li},
journal={Neurocomputing},
volume={338},
pages={139--153},
year={2019},
doi={10.1016/j.neucom.2019.01.036}
}
@article{liu2019roadnet,
title={RoadNet: Learning to Comprehensively Analyze Road Networks in Complex Urban Scenes from High-Resolution Remotely Sensed Images},
author={Liu, Yahui and Yao, Jian and Lu, Xiaohu and Xia, Menghan and Wang, Xingbo and Liu, Yuan},
journal={IEEE Transactions on Geoscience and Remote Sensing},
volume={57},
number={4},
pages={2043--2056},
year={2019},
doi={10.1109/TGRS.2018.2870871}
}
If you have any questions, please contact me without hesitation (yahui.liu AT unitn.it).