ciupava / fair-utilities Goto Github PK

Summary

With the aim to improve fAIr's performance in detecting buildings footprint, it is foremost needed to generate a metric that allows to understand how the tool performs. Subsequently, steps can be taken to improve the tool, and they will be assessed against the same metric.

Impact & value

Impact score: 10/10
Urgency score: 10/10
Difficulty: 8/10

What needs to be done?

• Run training + prediction on different datasets (e.g. another city/region of the world)
  • download datasets from OpenAerialMap and generate tiles in the same format of data/sample_2 (using this code?)
  • download correspondent labels from OSM data ...how?
  • run existing ramp code and generate new checkpoints for different cities
  • chose cities for comparison
  • perform different statistics/metrics to understand how well the tool does
• test the different zoom levels, to see impact (TBC)
• test different structures, either:
  • act on the existing RAMP architecture (turn on/off some layers)
  • scope for other TensorFlow architectures (find from what's available)

Who can help?

• Computer vision and Remote sensing experts
• @sgreenbury @yongrenjie @kshitijrajsharma @omranlm

---

Effort estimate

• 2/3 months time

Updates