Table 2.

Class-specific Model Performance on the Image Test Set for Three E-cigarette Product Detection Scenarios

Object class	PPV at 80% Sensitivity	PPV at 50% Sensitivity	PPV at 20% Sensitivity
Mod device	49%	73%	81%
Pod device	72%	86%	89%
E-juice container	97%	97%	99%
Packaging box	94%	98%	100%
Nicotine warning label	97%	100%	100%
E-juice flavor	91%	97%	100%
E-cigarette brand name	85%	97%	98%
Smoke cloud	38%	82%	95%

PPV = positive predictive value (also known as precision). Sensitivity is also known as “recall.” Bounding box overlap threshold (IoU) = 50%.