Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('csv', {}), NamedSplit('test'): ('csv', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

BONAI Dataset

Dataset Description

Dataset Summary

BONAI (Buildings in Off-Nadir Aerial Images) is a large-scale dataset for building footprint extraction (BFE) in off-nadir aerial images. It is the official dataset of the paper Learning to Extract Building Footprints from Off-Nadir Aerial Images (TPAMI 2022).

Unlike existing BFE datasets that only annotate building footprints, BONAI provides fully annotated instance-level roof and footprint polygons for each building, plus the corresponding offset vector from roof to footprint. This makes it possible to study the roof-to-footprint offset modeling that is characteristic of off-nadir imagery.

  • 268,958 building instances across 3,300 aerial images (1024Γ—1024)
  • Images collected from six representative cities in China: Shanghai, Beijing, Harbin, Jinan, Chengdu, Xi'an
  • Official official_v1 manifests and COCO annotations define 2,700 training, 300 validation, and 300 test images. Training/validation images are distributed together in trainval.zip.

Supported Tasks and Leaderboards

  • Building footprint extraction: given an off-nadir aerial image, predict the footprint polygon of each building instance (via roof segmentation + offset regression, or direct footprint prediction).
  • Building roof segmentation / object detection: single-class (building) instance segmentation and detection are directly supported by the COCO-format annotations.

The official evaluation code is tools/bonai/bonai_evaluation.py in the GitHub repository (requires the bstool library).

Languages

Annotations and documentation are in English. Image file names encode the source city and tile location.

Dataset Structure

Data Instances

Each image is a 1024Γ—1024 PNG aerial image tile. Each building instance carries:

  • roof polygon (segmentation + bbox)
  • footprint polygon (footprint_mask + footprint_bbox)
  • offset: 2D vector from roof to footprint, e.g. [-1, -4]
  • building_height, ignore_flag, area, bbox

Data Fields

The coco/*.json files follow the COCO format with BONAI-specific extensions:

Field Description
images id, file_name, height, width (1024Γ—1024), date_captured
annotations.segmentation roof polygon
annotations.roof_bbox roof bounding box [x, y, w, h]
annotations.footprint_mask footprint polygon
annotations.footprint_bbox footprint bounding box [x, y, w, h]
annotations.building_bbox enclosing building bounding box
annotations.offset roof→footprint offset vector [dx, dy]
annotations.building_height building height (0 if unknown)
annotations.ignore_flag 1 for ignored instances
annotations.area, bbox, iscrowd, image_id, category_id standard COCO fields
categories single class: building (id 1)

The csv/ files provide test-set polygons in WKT format with columns ImageId, BuildingId, PolygonWKT_Pix, Confidence (footprint and roof versions). Both versions cover the same 300 images as coco/bonai_shanghai_xian_test.json. The val in the legacy CSV filenames is a historical name; these files do not describe the validation subset. Use coco/bonai_val_official_v1.json for validation.

Data Splits

Official split: official_v1

Use the following image-level manifests and COCO annotations for the official training, validation, and test protocol:

Split Images Cities Image manifest COCO annotations
Train 2,700 Shanghai 1,656; Beijing 684; Chengdu 72; Harbin 288 train.txt bonai_train_official_v1.json
Validation 300 Shanghai 228; Jinan 72 val.txt bonai_val_official_v1.json
Test 300 Shanghai 200; Xi'an 100 test.txt bonai_shanghai_xian_test.json

The image counts match Table 1 of the paper. These downloadable manifests define official_v1; their identity with the historical image-level partition used for the paper results has not been established. Published paper metrics should not be treated as measured results for official_v1.

Geographic grouping. Shanghai has 157 parent tile groups, identified by the L18_<coordinate>_<coordinate> part of each filename. Every group contains 12 images: three image sources (arg, google, ms) and four 1024Γ—1024 crops per source. Using a fixed seed of 42, 19 complete groups are selected for validation (228 images), and the remaining 138 groups are used for training (1,656 images). All 72 Jinan images belong to validation. The Shanghai/Xi'an test set is unchanged.

No parent tile group crosses a split boundary. Checks of filename-derived parent tile extents find no positive-area overlap across train, validation, and test. Adjacent tile boundaries are allowed; no geographic buffer is imposed. This grouping check does not independently verify the imagery's georeferencing.

The split preserves every annotation field except COCO image and annotation IDs, which are made unique across the merged train/validation files. Original filenames are retained. The image ID map records each source file and original image ID; the protocol metadata records source hashes, selected groups, ID assignment, and raw annotation counts. Raw annotation counts include all released records and are not a claim of matching the paper's instance counts.

Shanghai geographic groups

Files and training configuration

trainval.zip continues to contain all 3,000 training/validation images. Both official train and validation COCO files reference those images; no separate image download or image renaming is needed. test.zip contains 300 image files and one directory entry (test/), not 301 images.

The default GitHub dataset configuration uses all five city-level trainval files for training, and its val entry also points to those same files. That is the 3,000-image combined-training protocol, not the independent validation protocol defined above. This configuration is also present in the March 2022 code release.

To train and validate with official_v1, download the two official COCO files into data/BONAI/coco/ in the BONAI code repository. Download loft_foa_r50_fpn_2x_bonai_official_v1.py into that repository's configs/loft_foa/ directory, beside the existing LOFT configuration. This override selects the separate train/validation COCO files and the existing test COCO file. The expected image directories are data/BONAI/trainval/images/ and data/BONAI/test/images/; arrange the extracted files accordingly.

python tools/train.py configs/loft_foa/loft_foa_r50_fpn_2x_bonai_official_v1.py

Use validation for development and model selection, and reserve test for final evaluation. The supplied configuration uses the repository's existing COCO evaluation hook for validation; the WKT files below remain test ground truth. The dedicated BONAI test entry point uses the independent Shanghai/Xi'an test set when passed --city shanghai_xian.

To regenerate the official manifests and COCO subsets from the original city-level annotations:

python scripts/create_official_split.py --data-root . --output-root .

The generation script uses complete parent tile groups and never changes the test set. The manifests are the authoritative membership lists for official_v1.

Test CSV filenames and compatibility

The following clearer filenames are byte-for-byte copies of the legacy files. The original files remain available because the existing GitHub evaluation script references their names.

Preferred filename in csv/ Legacy filename in csv/
bonai_shanghai_xian_test_footprint_minarea500.csv shanghai_xian_v3_merge_val_footprint_crop1024_gt_minarea500.csv
bonai_shanghai_xian_test_roof_minarea500.csv shanghai_xian_v3_merge_val_roof_crop1024_gt_minarea500.csv

All four files are test-set ground truth, not validation ground truth. The aliases do not add images or annotations. Original city-level trainval COCO files, image filenames, and archives remain available and unchanged.

File layout:

β”œβ”€β”€ trainval.zip
β”œβ”€β”€ test.zip
β”œβ”€β”€ coco/
β”‚   β”œβ”€β”€ bonai_{beijing,chengdu,haerbin,jinan,shanghai}_trainval.json
β”‚   β”œβ”€β”€ bonai_train_official_v1.json
β”‚   β”œβ”€β”€ bonai_val_official_v1.json
β”‚   └── bonai_shanghai_xian_test.json
β”œβ”€β”€ splits/official_v1/
β”‚   β”œβ”€β”€ train.txt
β”‚   β”œβ”€β”€ val.txt
β”‚   β”œβ”€β”€ test.txt
β”‚   β”œβ”€β”€ image_id_map.json
β”‚   β”œβ”€β”€ protocol.json
β”‚   └── shanghai_split.png
β”œβ”€β”€ scripts/create_official_split.py
β”œβ”€β”€ configs/loft_foa_r50_fpn_2x_bonai_official_v1.py
└── csv/                 # test ground truth; explicit test names and legacy aliases

Dataset Creation

Curation Rationale

Off-nadir aerial images exhibit visible roof-to-footprint displacement caused by the oblique viewing angle and building height. BONAI was curated to enable learning-based methods that explicitly model this offset, which nadir-only datasets cannot support.

Source Data

Aerial image tiles (1024Γ—1024) sourced from six Chinese cities β€” Shanghai, Beijing, Harbin, Jinan, Chengdu and Xi'an β€” covering diverse urban layouts and building styles. File names indicate the data source (e.g. shanghai_arg, shanghai_ms, shanghai_google).

Annotations

Instance-level annotations were produced for every building: roof polygon, footprint polygon and the offset vector between them, stored in COCO format with the extension fields listed above.

Personal and Sensitive Information

The dataset contains aerial imagery of urban areas. No personal or biometric information is annotated.

Considerations for Using the Data

Social Impact of Dataset

BONAI supports research in remote sensing and urban mapping (e.g. cadastral mapping, disaster assessment, 3D city modeling). Users should comply with local regulations when applying building extraction models.

Discussion of Biases

Imagery is limited to six Chinese cities; building styles, densities and imaging conditions may not generalize to other regions. Some instances are marked with ignore_flag where annotation is uncertain.

Other Known Limitations

  • building_height is 0 for instances where height is unknown.
  • The dataset targets building footprints only; other land-cover classes are not annotated.

Additional Information

Dataset Curators

Jinwang Wang, Lingxuan Meng, Weijia Li, Wen Yang, Lei Yu, Gui-Song Xia.

Licensing Information

The official code repository (jwwangchn/BONAI) is released under the MIT License. No separate license statement for the dataset files has been published; please refer to the repository and the paper authors for the intended terms of use.

Citation Information

@article{wang2022bonai,
  author={Wang, Jinwang and Meng, Lingxuan and Li, Weijia and Yang, Wen and Yu, Lei and Xia, Gui-Song},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title={Learning to Extract Building Footprints from Off-Nadir Aerial Images},
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2022.3162583}
}

Paper: arXiv Β· Code: jwwangchn/BONAI

Contributions

Contact: Jinwang Wang (jwwangchn@whu.edu.cn).

Downloads last month
1,327

Paper for jwwangchn/BONAI