Table 4 Evaluation of dimension reduction. For dogs, we use AlexNet and thus the dimension of full vector of this is 49,152
From: Refining deep convolutional features for improving fine-grained image recognition
Strategy | Cub (%) | Air (%) | Cars (%) | Dogs (%) | Dimension |
|---|---|---|---|---|---|
Full vector | 86.3 | 87.5 | 92.1 | 72.3 | 98,304/49,152 |
Tensor Sketch | 84.9 | 83.4 | 88.9 | 69.6 | 12,288 |
MI Selection 1 | 75.3 | 73.6 | 81.8 | 63.9 | 4,906 |
MI Selection 2 | 86.4 | 87.7 | 92.4 | 72.6 | 69,992/40,000 |
Tensor Sketch + MI | 84.5 | 82.5 | 87.5 | 68.4 | 4,906 |
PCA | 82.3 | 81.0 | 85.2 | 65.6 | 4,608 |