Swimming pool object detection
Uses AI algorithms to detect swimming pools.
Swimming Pool Object Detection recognizes swimming pools on 256x256 Pleiades imagery in GeoTIFF format.
This algorithm was trained on imagery in Argentina, it is currently not recommended to use this algorithm outside of South America.
- Pléiades Display (Streaming) -> Raster Tiling (tile size: 256) -> Swimming Pool Object Detection
- Pléiades Display (Download) -> DIMAP->GeoTIFF Conversion -> Raster Tiling (tile size: 256) -> Swimming Pool Object Detection
The algorithm uses RetinaNet network architecture, which allows improving the detection at different scales. It trains over thousands of swimming pools located in Pilar, BsAs. The input data consider different positions, shapes, colors, and background conditions, to cover all the possible circumstances. However, It also implements several tools of data argumentation to ensure the variability of the training data set.
|Input parameters||Pléiades Tile (256x256px)|
The algorithm reaches a mean average precision of 0.1649. The performance found to be strongly correlated with the scale of the object to the image size. Thus, the images were rescaled before been taken by the algorithm allowing the increase of the performance. An extra scale can be done by RetinaNet during the prediction (not the case during the training due to the computing cost). Also, the performance increase in open residential zones and decrease in urban commercial areas.
Algorithm Training Data Details
The algorithm trains over thousands of images which are split between 80% training and 20% validation. It trains with 50 epochs and 100 steps, it reaches a loss of 0.6 and mAP of 0.16.
Minimum Order Size
Note that the minimum AOI size for this processing block is 25 sqkm.
For more information about this processing block, please see the provider website.
Terms & Conditions
View the End User License Agreement conditions.