v2

Built by run 39829470042 from 9794832.

person.qnet is the int8 person detector majestic's CPU engine loads;
person.json, installed beside it, is what its three object-detection
presets mean for this model. Both the host engine and its ARMv7 NEON
build were checked bit for bit against the reference arithmetic before
anything was measured. Measured on labelled crops of COCO val2017:

3000 crops of COCO val2017 (1000 with nobody in them), 8743 people counted.

preset	threshold	precision	recall	F1	false alarms per person-free crop
fewerFalseAlarms	0.45	0.611	0.277	0.381	0.242
balanced	0.28	0.473	0.372	0.416	0.536
catchEverything	0.10	0.275	0.496	0.354	1.563

drum.qnet reads one drum of a utility meter's odometer: a 20x32 luma
crop in, 20 half steps of a digit out. Its host engine and ARMv7 NEON
build were checked against the reference bit for bit, input and every
logit, on 400 crops, and against each other byte for byte. What it was
trained on, and its attribution, are in models/drum/SOURCES.md.