Access to AVA Active Speaker Test Set

[Adrian Auer]

Hi, 

as I understand the AVA Active Speaker Challenge will not be hosted any longer and the last Event was in 2022, therefore I wanted to ask if it was possible to get access to the test set. Right now it is published only without the labels. We want to evaluate different models on it and measure their performances under different circumstance. Threfore we want to add more labels to the test set making it a Benachmark that also assesses the difficulty of different social situations. We want to give values for the videos under the following categories:

• how many people are in the video in total (Diversity)
• what is the max and average amount of people in the same frame (Interactivity)
• How much do the bounding boxes move on average (Dynamic)
• How much do bounding boxes overlap on average (Face Occlusion)
• How much speech is labeled NOT AUDIBLE in percentage of the whole speech (Audibility)
• How much noise is present (Noise) In the paper they also analyze for noise, but these labels are, as I understand, in the AVA Speech Dataset. 

Please let me know if there is the possibility to get access or are you planning to hold it back for any future challenges? 

Best, 
Adrian

[Sourish Chaudhuri]

Hi Adrian, thank you for your interest in doing more with the dataset, generally! I don't think we have any plans to organize future challenges with the data, and given that it's been multiple years since the last iteration of the public challenge, I don't see any issues with releasing the test data to have it be generally available. I'm curious if others have thoughts on this?

Thanks,

 Sourish

--
You received this message because you are subscribed to the Google Groups "AVA Dataset Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ava-dataset-us...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ava-dataset-users/44f988c7-bdd0-4c59-9f52-5a4e680f236en%40googlegroups.com.

[Adrian Auer]

Hi,

just wanted to follow up. Apparently nobody has any (other) thoughts on that. So how can we get access and do you need us to do somthing for that?

Cheers, 
Adrian