There is a significant error in the experimental results.

[Guoxn1]

After running the code, I found that the error in the attack results is larger than that in the paper. Specifically, under single frame attack conditions, the results of the model that was not attacked were consistent with the results in the paper, but the results after using a single frame attack were much better than those in the paper, such as fqa apollscape. The results I ran were about 90% higher than those in the paper

[chengqb0126]

After running the code, I found that the error in the attack results is larger than that in the paper. Specifically, under single frame attack conditions, the results of the model that was not attacked were consistent with the results in the paper, but the results after using a single frame attack were much better than those in the paper, such as fqa apollscape. The results I ran were about 90% higher than those in the paper在运行代码后，我发现攻击结果中的误差比论文中的要大。具体而言，在单帧攻击条件下，未被攻击的模型的结果与论文中的结果一致，但使用单帧攻击后的结果要比论文中的结果好得多，如fqa apollscape。我运行的结果比论文中的高出90%左右

兄弟，你跑了多久，我跑apolloscape+grip跑了几个小时还没跑完。