Normalizing Flow-Based Industrial Complex Background Anomaly Detection

Table 1

Comparison of area under ROC in % (AUROC) of different methods on MVTec-AD.

Method	STFPM	GANomaly	SPADE	PaDiM (R18-Rd100)	DifferNet	CS-Flow	CSA-Flow (ours)
Carpet	98.8	69.9	97.5	98.9	92.9	100.0	100.0
Grid	99.0	70.8	93.7	94.9	84.0	99.0	98.7
Leather	99.3	84.2	97.6	99.1	97.1	100.0	100.0
Tile	97.4	79.4	87.4	91.2	99.4	100.0	99.1
Wood	97.2	83.4	88.5	93.6	99.8	100.0	99.2
Texture classes	98.3	77.5	92.9	95.5	94.6	99.8	99.4
Bottle	98.8	89.2	98.4	98.1	99.0	99.8	99.8
Cable	95.5	75.7	97.2	95.8	95.9	99.1	98.7
Capsule	98.3	73.2	99.0	98.3	86.9	97.1	99.4
Hazelnut	98.5	78.5	99.1	97.7	99.3	99.6	100.0
Metal nut	97.6	70.0	98.1	96.7	96.1	99.1	98.1
Pill	97.8	74.3	96.5	94.7	88.8	98.6	97.5
Screw	98.3	74.6	98.9	97.4	96.3	97.6	96.9
Toothbrush	98.9	65.3	97.9	98.7	98.6	91.9	94.2
Transistor	82.5	79.2	94.1	97.2	91.1	99.3	99.5
Zipper	98.5	74.5	96.5	98.2	95.1	99.7	99.7
Object classes	96.5	75.5	97.6	97.3	94.7	98.2	98.4
Average	97.1	76.1	96.0	96.7	94.7	98.7	98.7

Bold values denote the best result in the category.