Skip to content

Model card

The full card lives in docs/model-card.md. Everything below is read from the report produced by the last run of ml/eval.py, so this page cannot show a number that was never measured.

Synthetic demo checkpoint. Trained on procedurally generated imagery of eight invented countries. Every metric below describes that synthetic world; none of it transfers to real-world geolocation.

Identity

Checkpoint
demo-convnext_atto-rig-h10
Model version
0.1.0
Dataset
synthetic-v1
Rig
rig-h10 (10 views)
Encoder
convnext_atto
Aggregator
attention

Held-out test metrics

56 sequences, geographically separated from training.

Top-1 accuracy
1.0000
Top-3 accuracy
1.0000
Macro F1
1.0000
Expected calibration error
0.5570
Brier score
0.3770
Temperature
1.0000

Calibration: not fitted: validation accuracy is 1.000, which leaves temperature scaling no errors to calibrate against. Temperature held at 1.0.

View-count ablation

The same checkpoint, evaluated with fewer views masked in.

1 view(s)
top-1 0.8393 · 79 ms
3 view(s)
top-1 0.9821 · 73 ms
5 view(s)
top-1 1.0000 · 69 ms
10 view(s)
top-1 1.0000 · 77 ms

Per country

CountryPrecisionRecallF1Support
XA1.0001.0001.0007
XB1.0001.0001.0007
XC1.0001.0001.0007
XD1.0001.0001.0007
XE1.0001.0001.0007
XF1.0001.0001.0007
XG1.0001.0001.0007
XH1.0001.0001.0007

Latency

Median per sequence (CPU)
68.2 ms

CPU only. This machine has a CPU-only torch build (2.14.0+cpu), so no GPU figure is reported rather than estimated.

Intended use

Demonstrating and evaluating multi-view visual geolocation. A prediction is probabilistic and describes the appearance of a place.

Out of scope

  • Identifying, tracking or inferring anything about a person.
  • Establishing someone’s nationality, residence or movements.
  • Any decision affecting a person’s rights, safety or access to services.
  • Operational or forensic geolocation of real imagery.