A classifier has high overall accuracy but fails frequently on one camera type. What analysis is most useful?

1. Measure only model download speed
2. Report only the global accuracy again
3. Remove every camera label from the evaluation metadata
4. Report performance by relevant data subgroups