GRAM Dataset
May 26, 2026 · View on GitHub
Labels, pseudo masks, and evaluation results for GRAM
This page provides the dataset used for GRAM (Generalized Region-Aware Mixture-of-Experts), presented at AAAI 2026. For details on the model and methodology, please refer to our paper.
⚠️ Note: Satellite imagery is not included due to licensing restrictions. Images can be accessed via ESRI World Imagery Wayback at zoom level 16 (~1.2 m/px). We used y_x as an image ID.
Source Dataset
Pseudo-label Dataset
Link: Google Drive
Pseudo-labels used for MoE source model training, generated by a semi-supervised segmentation model (For the code and dataset for the SSL segmentation model, refer to this paper). The number of pseudo-labels for each city is as follows. For the detail of dataset, please refer to the appendix of our paper.
| Continent | City | Total Grids |
|---|---|---|
| Africa | Cairo (Egypt) | 338,503 |
| Africa | Cape Town (South Africa) | 458,799 |
| Africa | Nairobi (Kenya) | 122,764 |
| Africa | Ouagadougou (Burkina Faso) | 147,546 |
| Asia | Colombo (Sri Lanka) | 96,339 |
| Asia | Karachi (Pakistan) | 321,073 |
| Asia | Mumbai (India) | 169,407 |
| South America | Caracas (Venezuela) | 97,232 |
| South America | Medellín (Colombia) | 83,256 |
| South America | Rio de Janeiro (Brazil) | 353,480 |
| Central America | Port-au-Prince (Haiti) | 29,286 |
| Central America | Tegucigalpa (Honduras) | 110,988 |
| Total | 2,328,673 |
Target Dataset
Ground-Truth Labels
Manually annotated GT labels for the 3 target cities.
Link: Source
| City | Labeled Grids | Total Grids |
|---|---|---|
| Dar es Salaam (Tanzania) | 6,050 | 237,974 |
| Kampala (Uganda) | 2,623 | 48,696 |
| Maputo (Mozambique) | 8,863 | 242,963 |
| Total | 17,536 | 529,633 |
Results
Evaluation results and performance (mIoU / F1-score) for target cities. Note that the spatial resolution of the released predictions has been reduced due to privacy considerations.
Link: Google Drive
| Method | Dar es Salaam | Kampala | Maputo |
|---|---|---|---|
| Vanilla Source | 0.681 / 0.792 | 0.716 / 0.814 | 0.800 / 0.888 |
| MoE Source | 0.806 / 0.885 | 0.800 / 0.881 | 0.900 / 0.947 |
| SHOT | 0.712 / 0.813 | 0.713 / 0.810 | 0.813 / 0.895 |
| TENT | 0.691 / 0.800 | 0.716 / 0.814 | 0.802 / 0.889 |
| CoTTA | 0.762 / 0.853 | 0.821 / 0.900 | 0.821 / 0.900 |
| SAR | 0.700 / 0.807 | 0.748 / 0.843 | 0.804 / 0.890 |
| BeCoTTA | 0.741 / 0.836 | 0.844 / 0.911 | 0.904 / 0.949 |
| GRAM (Ours) | 0.859 / 0.921 | 0.870 / 0.927 | 0.907 / 0.951 |
Citation
@inproceedings{lee2026gram,
title = {Generalizable Slum Detection from Satellite Imagery with Mixture-of-Experts},
author = {Lee, Sumin and Park, Sungwon and Yang, Jeasurk and Kim, Jihee and Cha, Meeyoung},
booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
year = {2026}
}