← Marketplace
Building Bounding Box Detector
Finds bounding boxes around clearly separated buildings.
Description
Building Bounding Box Detector finds bounding boxes that contain individual buildings. This block is faster, more accurate and significantly more affordable than traditional building detector approaches. It works best on Pléiades imagery with no limitations on image size.
Supported workflows
Data platform:
- Catalog -> Pléiades Display -> Order and get the asset from storage of UP42
- In Projects workflows, Processing from Storage -> DIMAP GeoTIFF conversion -> Building Bounding Box Detector
Data blocks:
- Pléiades Display (Streaming) -> Building Bounding Box Detector
Technical Description
It is intended for use on buildings that are clearly separated and countable. The output is provided as a JSON-file with the bounding box polygons across the entire area of interest which can be used in any other geospatial application.
Use cases range anywhere from cadastral mapping, humanitarian aid, urban sprawl monitoring, infrastructure/urban planning, damage assessment and much more.
| General Information | Description |
|---|---|
| Technical Information | The algorithm is a customized architecture that uses YOLO v3 as its primary building block. YOLO v3 itself out performs most other object detection algorithms (such as SSD and Faster-RCNN) in both speed and accuracy. |
| Performance | The algorithm has very high recall and pretty high precision on buildings that are more well separated. The model was also trained more on smaller more residential buildings. An example of were it would work well is a more suburbian residential area of LA. Examples of where it would not work well are downtown LA where there are lots of skyscrapers, or European old towns where buildings tend to be all connected and it’s hard to count even manually. Given he intended input, recall is roughly 90% with a precision of around 80% (mAP=0.7). |
| Required image resolution | Between 0.15 m and 0.6 m. |
| Input parameters | Imagery in GeoTIFF of any size. Requires 3 Channel RGB images. Likely target would be Pleiades Streaming imagery. |
| Output Format | A GeoJSON containing the bounding box polygons detected. |
Algorithm Training Data Details
Trained on the XView dataset. Additional reading can be found here:
- The XView Dataset and Baseline Results,
- On the importance of proper data handling (part 1)
- On the importance of proper data handling (part 2)
For more information, please visit the provider site.
Capabilities
Terms & Conditions
View the End User License Agreement conditions.