Datasets:

---

lpynow

/

MMA-82

like 8

Hi Junjie,

Thank you for reporting this issue.

Unfortunately, the two original YouTube links are no longer publicly available, and we are unable to redistribute the source videos directly due to potential copyright restrictions.

As an alternative, you may search for the same videos on the original content platform. The platform provides either a free trial or a paid membership option, which may allow you to access the full videos and reconstruct the corresponding clips, subject to the platform’s terms of service.

We will also look into whether we can provide updated source links or other legally accessible alternatives in a future update.

Thank you for your understanding!

Thank you for your previous reply, and I apologize for bothering you again.

I downloaded the CAER dataset, not CAER-S, from the official website:

https://caer-dataset.github.io/

However, when rebuilding the Emotion subset of MMA-82-Rec using mma82_release.py and mma82_release_metadata.json, I found that the crop intervals for many samples exceed the duration of the corresponding CAER source videos.

For example, E_Training000004.mp4 is mapped to:

source: train/Anger/0350.avi
start_time: 2.643182
end_time: 3.682993

However, the corresponding CAER source video is only about 1.53 seconds long, so the requested clip cannot be generated correctly.

Could you please confirm whether MMA-82 was built using a different version of CAER, or whether a specific CAER archive is required to rebuild the Emotion subset of MMA-82-Rec? If possible, could you also provide a download link or other information that would help identify the correct version of the CAER dataset?

Thank you again for your help.

Hi Junjie,

Thank you for your interest.

We have reviewed the dataset you requested. The dataset you need to download is CAER: https://drive.google.com/file/d/1JsdbBkulkIOqrchyDnML2GEmuwi6E_d2/view.

After checking, we found that the duration of train/anger/0350.avi is 3+ seconds. Please check to see if there was an error during the download.

Thank you for providing the download link.

I downloaded the CAER archive again from the Google Drive link you provided and checked train/Anger/0350.avi directly. However, the video in the downloaded archive still has the following properties:

Duration: 1.532618 seconds
Container: AVI (Audio Video Interleaved)
Codec: MPEG-4 Part 2
Resolution: 712 × 400
Pixel format: yuv420p
Frame count: 46
Frame rate: approximately 30 FPS
File size: 171,332 bytes
SHA256: 4fc1a6b103bcfac17ffad1906344c3b30bcfe70d8b9a26321fc104dd7be0fa7c

I have also uploaded the exact video file extracted from the downloaded archive here:

https://huggingface.co/datasets/zhou12345/CAER/resolve/main/train/Anger/0350.avi

Could you please confirm whether the CAER dataset available through this link is the same version that was used to construct MMA-82?

Thank you again for your help.

Upon review, we discovered that we had renamed the caer.zip file. We have uploaded the script to MMA82_Release/recreate_caer.py. You can use this script to recreate the caer dataset.

Thank you for providing recreate_caer.py.

However, the script requires the following mapping file:

comparison_reports/caer_to_caer1_mapping.csv

I could not find this file in the repository. Could you please also upload the mapping CSV?

We have uploaded the file to MMA82_Release