tsrobcvai's picture
Update README.md
a8a67ab verified
<h2 align="center">ROI-1555: Rebar Detection and Instance Segmentation Dataset</a></h3>
<div align="center">
<img src="doc/img1.png" width="900"></img>
</div>
* ROI-1555 for rebar object detection and instance segmentation contains 1555 rebar images and their fine-labeled bounding boxes and pixel-wise masks.
* Diverse rebar specifications, layouts, application scenarios, and environmental conditions.
## Usage
* Here is an example to convert the annotations to MSCOCO 2017 format
```
python
cp -r 1260/img_label tools/data_annotated/train2017
cd tools
python labelme2coco_instance.py
#Annotation instances_train2017.json(MSCOCO 2017 format) will be generated in tools/annotations
```
## Paper
<h2 align="center"><a href="https://www.sciencedirect.com/science/article/pii/S147403462500117X">Deep Learning-based Rebar Detection and Instance Segmentation in Images</a></h3>
<p align="center">
<a href="https://www.shao-lab.com/Team-c7cad4a2a33a4d7686e0b6e8a524b816">Tao Sun</a>,
Qipei Fan</a>,
<a href="https://www.shao-lab.com/Learn-more-4469738c5183485884effe68c04e692d">Yi Shao*</a>
<br>
Advanced Engineering Informatics
</p>
If you use the dataset in your work, please cite our paper:
```
@article{sun2025deep,
title={Deep learning-based rebar detection and instance segmentation in images},
author={Sun, Tao and Fan, Qipei and Shao, Yi},
journal={Advanced Engineering Informatics},
volume={65},
pages={103224},
year={2025},
publisher={Elsevier}
}
```
## Highlights
* Mask2Former trained on our dataset shows good generalization ability in unseen data
<div align="center">
<img src="doc/img2.png" width="900"></img>
</div>
* Benchmark test shows the performance and limitations of the popular networks
<div align="center">
<img src="doc/img3.png" width="900"></img>
</div>
* Six data augmentation strategies were introduced and tested to improve the SOTA method,
which can guide the selection of suitable data augmentation strategies for rebar perception.
<p align="center">
| Methods | Epochs | mAP | mAP50 | mAP75 | mAPstraight | mAPhoop |
|:--------------------------------:|:------:|:----:|:-----:|:-----:|:-----------:|:-------:|
| Without additional data augmentation | 178 | 68.6 | 93.6 | 79.9 | 76.0 | 61.1 |
| With Random Vertical-Flip | 178 | 60.1 | 91.7 | 65.8 | 70.1 | 50.1 |
| With Random Rotation | 178 | 66.9 | **94.4** | 77.6 | 75.1 | 58.8 |
| With Mosaic | 178 | 54.1 | 85.6 | 58.7 | 61.4 | 46.8 |
| With MixUp | 356 | 64.0 | **94.1** | 71.8 | 72.6 | 55.4 |
| With Cutout | 356 | **70.2** | **94.2** | **81.9** | **77.5** | **62.8** |
| With Simple Copy-Paste | 718 | **71.4** | **93.9** | **84.7** | **78.5** | **64.3** |
</p>
## Contact
If you have any questions on the dataset, please email [tao.sun@mail.mcgill.ca](mailto:tao.sun@mail.mcgill.ca).