# IEEE Bigdata Cup 2024: Building extraction — Data

来源：https://www.kaggle.com/competitions/building-extraction-generalization-2024/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

We present a comprehensive collection of datasets designed to encompass a broad spectrum of land use classifications in Japan. This collection includes a diverse selection of satellite imagery and detailed mappings of building footprints across a range of locales, including central, urban, suburban, and rural areas. The building footprint information provides precise local coordinates for the vertices of the building outline polygons, formatted in COCO standard (https://cocodataset.org/#home). 
The entire dataset is systematically organized into two parts: training set and test set. Figure 1 delineates the geographical coverage of the training and test sets, highlighting the concentration of the training set in the Tokyo vicinity, while the test sets are strategically dispersed across Japan. This distribution is designed to rigorously evaluate the model's generalization abilities.  Learning strategies such as transfer learning and domain adaptation are encouraged to use.
![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4876373%2F549622dd3280963b81ff4ed992521a6f%2F2.png?generation=1713961058316432&alt=media)
## Training dataset
The training set data uses 0.3-meter Google earth satellite images complemented by meticulously manually annotated building outlines. The training set data was collected from:
- **Susono**: Represents rural areas and contains 1,113 images, 18,159 building footprints
- **Hachioji**: Represents suburban areas and contains 1,234 images, 30,841 building footprints
- **Setagaya**: Represents urban areas and contains 1,272 images, 57,132 building footprints
- **Hachioji**: Represents downtown areas and contains 1,098 images, 63,979 building footprints

With a ratio of 8:2 the training data was divided into a training set and a validation set
## Test dataset
The imagery and building annotations for both test sets derive from the open-source Japanese 3D city model, the Plateau project (https://www.mlit.go.jp/plateau/) enhanced with manual adjustments following visual inspection. All test images were randomly selected from 42 cities in Japan, but a balance of different types of areas was maintained. A total of 250 images were taken from each region, totaling 1,000 images.


We assure that privacy concerns have been fully addressed in the compilation of this data. 

## Folders and files
- **train/** - a folder where the training data is stored. Where the image folder holds 3,784 training images and the train.json file records the annotations in coco format.
- **val/** - a folder where the validation data is stored. Where the image folder holds 933 validation images and the train.json file records the annotations in coco format.
- **test/** - the 1,000 images of the challenge test set.
- **sample_submission.csv** - a (truncated) sample submission file in the correct format. It includes 14 detections over (only) 6 images.

## Submission Format 
The extraction results of all images should to be saved in one single CSV file, with each line representing the results of one image. Each line must include the image id (***ImageID***) and a list of coordinates for the extraction results (***Coordinates***). The *ImageID* should match the image name but does not need to be four digits long (e.g., an image named 0001.jpg would have an *ImageID* of 1). 

The *Coordinates* column should contain a list of coordinates for the extracted mask for each building. These coordinates should be formatted in [(x0, y0), (x1, y1), (x2, y2), ... , (xn, yn)]. To facilitate reading the coordinate column with *Pandas* , the coordinates needs to be stored as a string (This means the square brackets [ ] should be enclosed in quotation marks " "). For detailed examples, refer to the ***sample_submission.csv***. 

If no building is extracted from a particular image, record an empty list ("[ ]" or []) in the *Coordinate* column instead of skipping the row. Therefore, your result file should contain 1000 lines, corresponding to the number of images in the test set.