Can you give the accurate number of few-shot accuracy linear prob?
chagmgang opened this issue · comments
Keumgang Cha commented
Can you give the accurate number of few-shot accuracy linear prob?
Delong Chen (陈德龙) commented
Hi, the followings are the few-shot evaluation result of RemoteCLIP in the Fig. 8 of our paper:
ResNet-50
n-shot | RSI-CB128 | RSI-CB256 | WHU-earth | EuroSAT | MLRSNet | PatternNet | RESISC45 | AID | RS2800 | OPTIMAL-31 | RSC11 | WHU-RS19 |
---|---|---|---|---|---|---|---|---|---|---|---|---|
1-shot | 35.59 | 42.52 | 44.25 | 43.20 | 31.75 | 46.10 | 39.33 | 36.95 | 41.93 | 42.80 | 48.13 | 45.15 |
4-shot | 60.04 | 65.44 | 62.96 | 55.53 | 46.90 | 66.99 | 52.11 | 63.13 | 56.21 | 62.63 | 61.67 | 73.59 |
8-shot | 69.55 | 75.89 | 75.38 | 61.75 | 55.02 | 77.07 | 61.75 | 70.50 | 63.29 | 73.01 | 72.51 | 85.44 |
16-shot | 77.58 | 83.72 | 77.67 | 70.36 | 59.74 | 82.93 | 69.51 | 75.12 | 73.86 | 75.38 | 76.57 | 89.32 |
32-shot | 82.02 | 87.04 | 85.08 | 77.44 | 64.99 | 88.32 | 75.71 | 82.46 | 77.71 | 81.61 | 83.82 | 93.79 |
ViT-B-32
n-shot | RSI-CB128 | RSI-CB256 | WHU-earth | EuroSAT | MLRSNet | PatternNet | RESISC45 | AID | RS2800 | OPTIMAL-31 | RSC11 | WHU-RS19 |
---|---|---|---|---|---|---|---|---|---|---|---|---|
1-shot | 34.31 | 44.28 | 41.25 | 44.89 | 34.14 | 45.98 | 42.10 | 37.04 | 39.29 | 41.13 | 44.14 | 40.78 |
4-shot | 64.49 | 70.33 | 60.67 | 55.99 | 54.52 | 70.98 | 60.91 | 65.59 | 55.96 | 69.30 | 61.99 | 68.16 |
8-shot | 76.13 | 83.73 | 75.71 | 65.76 | 64.24 | 82.53 | 70.92 | 75.72 | 63.75 | 77.85 | 72.43 | 80.68 |
16-shot | 82.63 | 89.12 | 77.50 | 75.73 | 67.45 | 88.13 | 75.83 | 81.05 | 71.61 | 82.20 | 77.13 | 89.51 |
32-shot | 88.11 | 91.83 | 87.08 | 83.30 | 71.58 | 91.87 | 81.77 | 86.67 | 81.61 | 89.14 | 85.50 | 93.40 |
Keumgang Cha commented
So, in the case of 32-shot, only 32 shots per class were used for learning and the rest were used for testing?
Delong Chen (陈德龙) commented
Actually no, we randomly sample different amount (1-shot to 32-shot) of few-shot examples from the 80% training split, and keep the testing set the same - all settings use the remaining 20% testing split.
Keumgang Cha commented
Thank you!