Skip to main content
. Author manuscript; available in PMC: 2023 Sep 18.
Published in final edited form as: IEEE Winter Conf Appl Comput Vis. 2023 Feb 6;2023:4709–4719. doi: 10.1109/wacv56688.2023.00470

Table 2.

OOD detection performance for different baselines. Near-OOD represents label shift, with samples from the unseen classes of the same dataset. Far-OOD samples are from a separate dataset. The numbers are averaged over five runs.

Train Dataset Method Near-OOD (Wild) Far-OOD (CIFAR10) Far-OOD (CelebA)
AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95
AFHQ Baseline 0.88±0.04 47.40±5.2 0.95±0.04 73.59±9.4 0.95±0.03 70.69±8.9
Baseline+TS [17] 0.88±0.03 45.53±9.8 0.95±0.04 71.77±8.9 0.95±0.03 65.89±8.3
Baseline+TS+ODIN [38] 0.87±0.05 45.02±1.51 0.95±0.05 69.42±2.38 0.95±0.03 67.18±2.16
Baseline+energy [42] 0.88±0.03 47.77±1.10 0.94±0.05 72.68±2.69 0.96±0.04 74.75±2.89
Mixup [70] 0.86±0.06 53.83±6.8 0.82±0.11 57.01±8.6 0.88±0.13 70.51±9.8
DUQ [65] 0.78±0.05 20.98±2.0 0.67±0.59 16.23±1.5 0.66±0.55 15.34±2.6
DDU [47] 0.83±0.02 23.19±2.6 0.90±0.02 32.98±10 0.75±0.02 10.32±5.6
Baseline+ACE 0.89±0.03 51.39±4.4 0.98±0.02 88.71±5.7 0.97±0.03 88.87±9.8

MC-Dropout [9] 0.84±0.09 30.78±2.9 0.94±0.02 73.59±2.1 0.95±0.02 71.23±1.9
5-Ensemble [32] 0.99±0.01 65.73±1.2 0.97±0.02 89.91±0.9 0.99±0.01 92.12±0.7
Near-OOD (Digits 7–9) Far-OOD (SVHN) Far-OOD (fMNIST)
AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95

Dirty MNIST Baseline 0.86±0.04 28.23±2.9 0.75±0.15 51.98±0.9 0.87±0.02 58.12±1.5
Baseline+TS [17] 0.86±0.01 30.12±2.1 0.73±0.07 48.12±1.5 0.89±0.01 61.71±2.8
Baseline+TS+ODIN [38] 0.83±0.04 34.13±12.07 0.77±0.13 21.59±19.62 0.89±0.02 46.43±4.31
Baseline+energy [42] 0.87±0.04 40.30±1.05 0.86±0.12 43.92±2.30 0.91±0.02 62.10±5.17
Mixup [70] 0.86±0.02 35.46±1.0 0.95±0.03 65.12±3.1 0.94±0.05 66.00±0.8
DUQ [65] 0.78±0.01 15.26±3.9 0.73±0.03 45.23±1.9 0.75±0.03 50.29±3.1
DDU [47] 0.67±0.07 10.23±0.9 0.68±0.04 39.31±2.2 0.85±0.02 53.76±3.7
Baseline+ACE 0.94±0.02 37.23±1.9 0.98±0.02 67.88±3.1 0.97±0.02 70.71±1.1

MC-Dropout [9] 0.97±0.02 40.89±1.5 0.95±0.01 62.12±5.7 0.93±0.02 65.01±0.7
5-Ensemble [32] 0.98±0.02 42.17±1.0 0.82±0.03 55.12±2.1 0.94±0.01 64.19±4.2
Near-OOD (Kids) Far-OOD (AFHQ) Far-OOD (CIFAR10)
AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95

CelebA Baseline 0.84±0.02 1.25±0.1 0.86±0.03 88.57±0.9 0.79±0.02 29.01±5.1
Baseline+TS [17] 0.82±0.04 1.24±0.1 0.87±0.06 88.75±0.9 0.78±0.04 29.01±5.1
Baseline+TS+ODIN [38] 0.65±0.01 8.75±2.21 0.55±0.01 23.03±0.16 0.54±0.01 5.00±0.07
Baseline+energy [42] 0.76±0.51 9.40±0.01 0.94±0.08 32.08±1.70 0.85±0.76 17.10±0.72
Mixup [70] 0.82±0.08 22.18±2.7 0.95±0.02 82.96±2.5 0.79±0.13 30.54±1.3
DUQ [65] 0.80±0.03 14.68±3.1 0.72±0.07 26.62±7.7 0.86±0.04 28.70±4.1
DDU [47] 0.73±0.15 7.9±0.4 0.74±0.13 8.18±0.4 0.81±0.15 25.45±1.4
Baseline+ACE 0.87±0.03 34.37±2.5 0.96±0.01 96.35±2.5 0.92±0.05 63.51±1.5

MC-Dropout [9] 0.70±0.10 25.62±1.7 0.86±0.1 91.72±7.5 0.74±0.12 64.79±1.8
5-Ensemble [32] 0.93±0.03 10.35±0.2 0.99±0.0 98.31±1.2 0.92±0.10 61.88±1.2
Near-OOD (other lesions) Far-OOD (CelebA) Far-OOD (Skin-texture)
AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95 AUC-ROC TNR@TPR95

Skin Lesion Baseline 0.67±0.04 8.70±2.5 0.66±0.06 10.00±3.6 0.65±0.10 5.91±2.8
Baseline+TS [17] 0.67±0.05 8.69±2.0 0.63±0.06 9.24±4.3 0.68±0.07 5.70±3.2
Baseline+TS+ODIN [38] 0.68±0.01 9.43±0.33 0.67±0.07 11.32±4.66 0.68±0.07 6.60±0.29
Baseline+energy [42] 0.70±0.04 10.85±0.08 0.70±0.14 7.90±0.29 0.65±0.20 2.83±1.33
Mixup [70] 0.67±0.01 8.52±2.8 0.64±0.08 10.21±4.0 0.72±0.05 5.26±3.1
DUQ [65] 0.67±0.04 3.12±1.8 0.89±0.09 11.89±2.5 0.64±0.03 4.8±1.5
DDU [47] 0.65±0.03 3.45±1.9 0.75±0.04 15.45±2.9 0.71±0.05 4.19±1.3
Baseline+ACE 0.72±0.04 10.99±2.8 0.97±0.02 66.77±1.4 0.96±0.03 95.83±5.0

MC-Dropout [9] 0.67±0.05 9.45±3.9 0.80±0.07 30.00±3.2 0.56±0.03 10.87±2.3
5-Ensemble [32] 0.88±0.01 11.23±1.7 0.91±0.03 27.89±5.9 0.76±0.02 17.89±3.5