Download
Programs, code, visuals, and data for downloading. Some are shared through GitHub and ObservableHQ.
Fast and Flexible Cartogram (equal-density map) Production
F4Carto, the recommended new version
Installer for Windows 64 bit Program
Zipped package, no installation needed
Documentation
Carto3F, the older version
GeoAI Models for Object Detection in Urban Environment
Supported by PSC-CUNY, Center for Advanced Research on Spatial Information (CARSI, led by Prof. Sean Ahearn), and the MS in GeoInformatics (MGEOi) program, this project explores the application of advanced deep learning models for detecting small urban objects using high-resolution (6-inch) aerial imagery.
The current training phase is exploratory, utilizing a limited subset of the full dataset—merely 500 images—and 100 training epochs. Each image consists of a 1024 × 1024 pixels with oriented bounding boxes for three target feature types: water tanks, cooling towers/AC units, and curb cuts.
Six customized models have been successfully implemented and tested across two leading frameworks—MMDetection and Ultralytics YOLO—within multiple Python computational environments (Windows, macOS, and WSL/Ubuntu). The models demonstrate flexibility and compatibility across diverse GPU configurations, providing a strong foundation for future large-scale experimentation and refinement.
The preliminary results are as follows, and pre-trained PyTorch models that can predict these three features are available upon request.
| Base Model | GPUs | mAP(50) | |||
|---|---|---|---|---|---|
| Overall | Water Tank | Cooling Tower | Curb Cut | ||
| Rotated Faster R-CNN * | 1/CUDA | .545 | .794 | .551 | .290 |
| RoI (Regions of Interest) Transformer | 1/CUDA | .595 | .866 | .622 | .296 |
| LSKNet (Large Selective Kernel) * | 1/CUDA | .651 | .898 | .662 | .393 |
| Strip R-CNN | 1/CUDA | .632 | .867 | .642 | .385 |
| YOLO v11 | 1/CUDA | .787 | .948 | .798 | .616 |
| YOLO v12† | M3 Ultra/MPS | .645 | .929 | .635 | .370 |
| YOLO v12‡ | 3/CUDA | .766 | .941 | .756 | .602 |
*: Models trained with MMDetection and MMRotate frameworks by applying transfer learning techniques.
†: This Yolo v12 model does not use any pre-trained parameters like other models. It was trained using 500 NYC images with 1,600 epochs. Other models use parameters pre-trained from the DOTA V1 dataset, by default.
‡: This model uses pre-trained model parameters that are based on 5,000 NYC images of 1024 by 1024. The pre-training was accomplished by this project with a self-supervised learning (SSL) framework.