Yuecong Min

Yes，as mentioned in our paper, we adopted three kinds of data augmentation strategies (random crop, horizontal flip and random temporal scaling) during training.

Thanks for your attention, If your resolusion of video data is pretty high, perhaps a human detection can preserve more useful information before resizing the whole image. Our recent version...

Similar issue to #13.

It seems that you only train the baseline without the proposed VAC or SMKD, please follow the [Readme.md](https://github.com/ycmin95/VAC_CSLR/blob/71f3e0334fbc8cecc7ce9816ec69781068abaac0/README.md?plain=1#L36) to set the configuration file. We remove the evaluation on the test...

Hi, @kido1412y2y Can you report the evaluation results with a batch size larger than 1? I never run experiments with batch size of 1, and not sure the influence of...

@neccam if it is not convenient for releasing previous work, can you release the translation results of G2T and S2G2T models?

In our experiments, it will produce the deterministic result with the same random seed. As you can find in the [code](https://github.com/ycmin95/VAC_CSLR/blob/main/utils/random_state.py#L25), we fix most of random seeds.

It may depend on the environment and the device, thus I couldn't provide clear answer to this question.

> It may depend on the environment and the device, thus I couldn't provide clear answer to this question. For more information: [REPRODUCIBILITY](https://pytorch.org/docs/stable/notes/randomness.html)

Perhaps you can check whether the sentence is decoded successfully and the groudtruth file is configured as wished.