YOLOv11-Sports-Facilities

Geo Data(shapefile) Mining using YOLOv11-obb + SAHI

v1.0.0 Updated Jan 3, 2025

---

Model

In the updated repository, we implemented the latest Ultralytics YOLOv11-obb to learn our outdoor sports facilities dataset. On our validation dataset, YOLOv11n-obb and YOLOv11s-obb achieved very high mAP:

Model	Input Size	Precision(B)	Recall(B)	mAP50(B)	mAP50-95(B)
yolov11n-obb-100	1024 * 1024px	0.9647	0.93973	0.97289	0.87496
yolov11n-obb-150	1024 * 1024px	0.97194	0.92553	0.96654	0.89414
yolov11s-obb-200	1024 * 1024px	0.96612	0.95119	0.98045	0.93171

We found that although a larger model cannot get improvement in P & R, the detection boxes are much more satisfying.

Note that yolov11s-obb-200 may be over-fitted.

Interference and Shapefile

We found that using SAHI(Structured Attention-based Hierarchical Inference) can greatly improve the performance of the pre-trained model on 2048 * 2048px images. You can find both hbb box and obb mask in the picture below

Also, we found that the detection results of the obb model in SAHI are the sahi.annotation.mask object. So we provide a pipeline of how to draw the mask directly into the ESRI polygon, which is amazingly precise.

Here is a demo of our shapefile in ArcGIS Pro:

You may try more images in the test_images folder, they are actually level 19 images of Nanjing, 2024. So you can feel the strong transfer ability of YOLOv11-obb.

Dataset from our lab

Sincere appreciation to the original contributor, Mr. Liu Chang who once worked in the Key Laboratory of Geographic Information Science, Ministry of Education, East China Normal University, Shanghai, People’s Republic of China.
Our further plan is to enrich the dataset into a benchmark set of outdoor sports facilities in the Yangtze River Delta region.

Model choices

This dataset uses a label style consisting of DOTA obb style like:

x1	y1	x2	y2	x3	y3	x4	y4	classname	difficult
1686.0	1517.0	1695.0	1511.0	1711.0	1535.0	1700.0	1541.0	large-vehicle	1

So it can be used in any object detection model that accepts the DOTA obb style.

In the competition, we implemented the repository below:
https://github.com/hukaixuan19970627/yolov5_obb & https://github.com/SSTato/YOLOv7_obb
Sincere appreciation and best wishes to the writers, 胡凯旋 & SSTato!

In release v1.0.0, we provide a new version of over dataset for YOLOv11-obb, in such a structure:

dataset/
├── images/
│   ├── train/
│   │   └── 1024px*1024px images
│   └── val/
│       └── 1024px*1024px images
└── labels/
    ├── train_original/
    │   └── ...(DOTA labels)
    ├── val_original/
    │   └── ...(DOTA labels)
    ├── train/
    │   └── ...(YOLO-obb labels)
    └── val/
        └── ...(YOLO-obb labels)

So now, our dataset can be used directly in Ultralytics YOLO-obb in this style:

class_index	x1	y1	x2	y2	x3	y3	x4	y4
0	0.780811	0.743961	0.782371	0.74686	0.777691	0.752174	0.776131	0.749758

Note that the coordinates used in yolo-obb must be standardized. We provide the transform function in the notebook.

Code Implementation

In yolov11-obb.ipynb we provide the whole pipeline from transform dota labels to yolo style -> train -> Predict with SAHI -> Save detection results to ESRI Shapefile.

The only requirements are PyTorch, Ultralytics, SAHI, and a GPU with no less than 8GB of CUDA memory.

---

v1.0.0 Updated Jan 3 2025