Table 4.

Comparison of the ESS-MB trained on House100K with various augmentation settings

Pretext dataset	Augmentation	Pretext task	Downstream ImageNet classification
		Training loss ↓	Training loss ↓	Test loss ↓	Test accuracy (%) ↑
House100K	–	4.08 ± 0.003	6.13 ± 0.229	6.20 ± 0.18	9.70 ± 0.16
House100KLighting	–	4.07 ± 0.005	5.77 ± 0.314	5.92 ± 0.35	14.09 ± 0.25
House100K	✓	4.00 ± 0.005	4.67 ± 0.002	4.75 ± 0.01	18.05 ± 0.04
House100KLighting	✓	4.03 ± 0.001	4.49 ± 0.013	4.51 ± 0.01	20.74 ± 0.17∗

The column “augmentation” indicates whether the pretext training uses the augmentation method from the original MoCo. The best downstream classification result for the datasets is indicated with an asterisk.