zhipeng-wei / Image-to-Video-I2V-attack

Cross-Modal Transferable Adversarial Attacks from Images to Videos (CVPR 2022)
15 stars 3 forks source link

About UCF-101 fine-tuned model #8

Open SignedQiu opened 9 months ago

SignedQiu commented 9 months ago

Upon reviewing the reference_ucf101.py, I observed that the UCF-101 models showcased subpar performance on the UCF-101 dataset, particularly when applying varied data augmentation techniques. Further examination of the code revealed that the fine-tuned UCF-101 models, curiously, are configured with 400 classes, which is perplexing. This raises questions regarding the rationale behind such a configuration and its impact on the models' vulnerability to attacks.

zhangjl919 commented 8 months ago

前辈,可以一起讨论吗

zhipeng-wei commented 8 months ago

UCF-101 configuration file is not provided in the package gluoncv. Therefore, we leverage the Kinetics-400 configuration file to load UCF-101 models. We don't evaluate the impact of these configuration paramters in the adversarial vulnerability. It may be a future research direction to explore.