CMU-Mosei datasets

[yanzichuan]

I hope this message finds you well. Thank you for your excellent work on MARLIN and for sharing your valuable resources with the research community.

I am currently working on a project involving the CMU-MOSEI dataset and noticed that your work utilizes this dataset. I would like to ask if you could kindly provide the official download link or any guidance on where to access the CMU-MOSEI dataset.

Your support would be greatly appreciated, and I look forward to learning from your work. Thank you for your time and assistance.

[ControlNet]

I just checked their github and see their dataset looks not available currently. ecfm/CMU-MultimodalSDK#1

When downloading the dataset, you can find a csd file CMU_MOSEI_Labels.csd, and the Raw/Videos/Full/Combined/*.mp4 video files. These video files are processed with FaceXZoo to extract the facial video, then run MARLIN on it.

[yanzichuan]

I have reproduced this MOSEI accuracy rate is only 75 %, very accurate on celebv, but why is it not accurate on the CMU-Mosei dataset ? Do you need to modify what configuration On cmumosei, the accuracy of each task is about 10 points worse. Please also guide you

[cmxlucky]

I just checked their github and see their dataset looks not available currently. ecfm/CMU-MultimodalSDK#1我刚查了他们的 GitHub，发现他们的数据集目前似乎无法使用。 ecfm/CMU-MultimodalSDK#1

When downloading the dataset, you can find a csd file CMU_MOSEI_Labels.csd, and the Raw/Videos/Full/Combined/*.mp4 video files. These video files are processed with FaceXZoo to extract the facial video, then run MARLIN on it.下载数据集时，你可以找到一个 csd 文件 CMU_MOSEI_Labels.csd，还有 Raw/Videos/Full/Combined/*.mp4 视频文件。这些视频文件会用 FaceXZoo 处理提取面部视频，然后运行 MARLIN。

Could you clarify how Acc2 and Acc7 are evaluated on the MOSEI dataset in the paper? In particular, for Acc2, is the split based on > 0 vs < 0, or ≥ 0 vs < 0?