Training/Validation Data Split about disco

[Wangt-CN]

Hi @Boese0601 , thanks for your nice comment.

For training, we use 000-334 of TikTok dataset; For testing, we find that there are potential risks of the person ID leakage for the TikTok dataset. Therefore, we choose to collect 10 short videos from both the 335-340 sequence and the Internet to make sure there are no person ID coincide for fair comparison.

[Boese0601]

Thanks for your kind reply! That makes things clear. Btw could you please also upload those additional video sequences collected from the internet to Google Drive?

[Wangt-CN]

Hi @Boese0601 , I have submitted the query to the corporation to open-source the additional TikTok-style data. Since it is collected by the corporation so we need to get the permission.
Currently, if you want to make a fair comparison, you could follow the penultimate line of Table 1 which does not use the additional data for training.

[notorious-eric]

Hi, I download the tsv file and found that there are additional data in the file. Therefore, in the penultimate line of Table 1, you do not use the tsv file you presented, just use 335-340 sequence for evaluation, is that correct?

[Wangt-CN]

@notorious-eric Hi, do you mean the evaluation data? All the models are evaluated on the same data, i.e., 10 videos which is the combination of the original testing tiktok and additional data.

[Kelu007]

Hi @Boese0601 , thanks for your nice comment.

For training, we use 000-334 of TikTok dataset; For testing, we find that there are potential risks of the person ID leakage for the TikTok dataset. Therefore, we choose to collect 10 short videos from both the 335-340 sequence and the Internet to make sure there are no person ID coincide for fair comparison.

What are the videos collected from the Internet for evaluation?