thank you for your release about cvpr2022-ftcl HOT 4 CLOSED

1. Zero. However, the same random seed will also lead to different results on different machines.
2. Yes, according to my experience, it is enough in most cases. The performance improvement after epoch 500 is marginal.
3. Sorry, I do not fully understand what you mean. After the training process ends, changing the random number seed will not affect the test results. However, you can try different random seeds to obtain better ACM-Net model. According to our experiments, the training of ACM-Net is not very stable with different random seeds. Therefore, before training the FTCL network, it should be guaranteed that ACM-Net can achieve the result (42.6% mAP@Avg) reported in its paper under the current random seed.

from cvpr2022-ftcl.

thank you! I almost have the same environment as you mentioned, but it seems the pretrained result is unpredictable,
if the pretrained ACMnet cannot reach 42.6,is it possible to reproduce FTCL result?

from cvpr2022-ftcl.

I think that might be difficult to achieve. The results of FTCL are obtained when the pre-trained ACM-Net reaches 42.6. FTCL can generally improve the performance of existing methods, but its final result is still relevant to the backbone's performance. The discussions in the repository of ACM-Net might be helpful for you to reproduce their results.

from cvpr2022-ftcl.

thanks

from cvpr2022-ftcl.