search

raoyongming / CAL

[ICCV 2021] Counterfactual Attention Learning for Fine-Grained Visual Categorization and Re-identification
MIT License
145 stars 25 forks source link

为什么没用和WS-DAN使用相同的主干网络？ #2

Closed Stephen-Hao closed 3 years ago

Stephen-Hao commented 3 years ago

WS-DAN 使用的主干网络是inceptionV3，而CAL使用的是resnet101,后者性能明显比前者好，为什么论文中在不同数据集上对比时，没用也使用inceptionV3呢

raoyongming commented 3 years ago

Thanks for your interest in our paper.

We use the Pytorch implementation of WS-DAN from this repo. The performance is lower than the original paper (88.28 vs. 89.4). To better compare with SOTA methods like API-NET, we use a stronger backbone.

In our early experiments, we also found CAL can also improve WS-DAN based on Inception V3 (88.28 -> 89.32 on CUB).

Stephen-Hao commented 3 years ago

@raoyongming 好的，可能是我没有仔细看论文，感谢您的解答。