Open
Description
Recently, I have updated the performance table to follow the recent report format. I reviewed 10 recent papers as follows, and found that error rates and 10-crop testing are no longer used for evaluation after about 2017.
| measure | 10-crop | |
|---|---|---|
| ResNeXt (CVPR 2017) | err | O |
| DenseNet (CVPR 2017) | err | O |
| PolyNet (CVPR 2017) | err | O |
| DPN (NIPS 2017) | err | - |
| SENet (CVPR 2018) | err | - |
| MobileNetV2 (CVPR 2018) | acc | - |
| NASNet (CVPR 2018) | acc | - |
| PNASNet (ECCV 2018) | acc | - |
| AmoebaNet (AAAI 2019) | acc | - |
| EfficientNet (ICML 2019) | acc | - |
For the above reason, the recent commit replaces the top-k error rates with accuracies and removes the 10-5 column (the top-5 score with 10 crops).