What you're looking at: A real 0.5m Sky-View Factor image from the Chactun
archaeological zone, Calakmul, Mexico. Vegetation has been digitally stripped away.
Ancient platforms, buildings, and walls appear as geometric shadows and raised edges.
Paint the grid — click or drag over cells you think contain archaeological structures.
Don't overthink it: cover any area that looks geometric or man-made.
Then submit to see how you compare to the ML model.
Data: Kokalj et al. 2023, CC BY 4.0. Tile 1687, Chactun area.
You vs. the model
|
You |
ML Model |
| Cells correct (TP) | — | — |
| False alarms (FP) | — | — |
| Missed (FN) | — | — |
| Precision | — | — |
| Recall | — | — |
Note: scores above use 64 grid cells (60 px each). F1 scores in the lecture — 0.80 for Character et al., 0.89 for Britton et al. — are pixel-level across hundreds of square kilometers. Same concept, very different resolution.
Your marks vs. ground truth
Model predictions vs. ground truth
Correct (structure found)
False alarm (nothing there)
Missed structure
The Model Output page uses this same model — explore the full
prediction heatmap and ground truth side by side across multiple tiles.
The Threshold Lab then shows how adjusting the confidence
cutoff shifts the balance between precision and recall.