CMU-MultiComp-Lab/CMU-MultimodalSDK

Question about MOSI dataset

mishimario opened this issue · 5 comments

Hi I have a question about MOSI dataset.
When I checked other research codes on MOSI data, I found they use vision features whose final dimension is 20 and audio features whose final dimension is 5. But I think those features can't be downloaded from this repo.
Do you know something?
Thank you in advance!

I would suggest contacting the authors. It's possible they used an earlier version of the dataset, but most likely they are using a subset of the dataset features or summative features (e.g., mean, median)

Thank you for replying! Yes they might use a subset of the dataset. But a couple of papers are using those features. So I thought they used some different version of the dataset. Was there an earlier version of MOSI datset? And do you know it had some option of encoding audio and vision features into 5 and 20 features? I guess this repo was created lately? So I guessed some version of dataset was gone.

And I downloaded MOSI dataset from this repo. But the number of data is 2183 (train: 1283, valid: 214, test: 686).
According to those papers, it should be 2199(train: 1284, valid: 229, test : 686). Do you know something about it?

Thank you for replying! Yes they might use a subset of the dataset. But a couple of papers are using those features. So I thought they used some different version of the dataset. Was there an earlier version of MOSI datset? And do you know it had some option of encoding audio and vision features into 5 and 20 features? I guess this repo was created lately? So I guessed some version of dataset was gone.
Hello, I also encountered this problem. Have you solved it?

@Katyawa @mishimario I also encountered this problem. Have you solved it?