Classification (Pretrained on ImageNet)¶
We provide benchmarks of different finetune algorithms on CUB-200-2011, StanfordCars, Aircraft, StanfordDogs, Oxford-III-Pet and COCO-70 .
Those domain adaptation algorithms includes:
Note
We found that StanfordDogs, Oxford-III-Pet and COCO-70 have similar categories as ImageNet, thus most fine-tune algorithms take little effect on those datasets. Therefore we do not report the results.
CUB-200-2011 accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
46.7 |
59.6 |
70.3 |
79.0 |
BSS |
49.3 |
62.8 |
72.4 |
79.9 |
DELTA |
51.1 |
64.1 |
73.7 |
80.5 |
StochNorm |
51.6 |
63.5 |
72.7 |
80.4 |
Co-Tuning |
53.8 |
66.6 |
74.9 |
81.2 |
Stanford Cars accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
40.0 |
63.8 |
76.8 |
87.7 |
BSS |
43.5 |
67.5 |
78.3 |
88.0 |
DELTA |
44.3 |
67.9 |
79.8 |
88.3 |
StochNorm |
43.8 |
68.2 |
79.3 |
88.0 |
Co-Tuning |
48.8 |
71.6 |
82.0 |
89.2 |
FGVC Aircraft accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
42.6 |
60.6 |
70.4 |
81.9 |
BSS |
44.2 |
62.3 |
71.1 |
82.1 |
DELTA |
46.4 |
63.2 |
71.7 |
82.3 |
StochNorm |
46.7 |
63.2 |
71.7 |
81.9 |
Co-Tuning |
45.7 |
62.3 |
72.5 |
83.0 |
Stanford Dogs accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
82.7 |
85.3 |
86.5 |
87.3 |
DELTA |
84.5 |
86.4 |
87.3 |
88.2 |
Oxford-III Pet accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
89.1 |
90.9 |
91.7 |
93.1 |
DELTA |
90.6 |
91.9 |
92.8 |
93.7 |
COCO-70 accuracy on ResNet-50¶
Methods |
15% |
30% |
50% |
100% |
Baseline |
77.3 |
80.2 |
82.6 |
84.4 |
DELTA |
79.2 |
81.7 |
83.5 |
84.6 |
