Panowan dataset details

[batman007herewego]

Hello, thank you for compiling the approximately 12k Panowan dataset. I noticed that the Panowan 360 and 360+x datasets are all segmented into 10-second clips, and the provided YouTube links are also segmented into 10-second formats. Are there any other datasets that include the 944h duration mentioned in the paper (12k * 10s ≈ 35h)? Thank you.

[yousiki]

Thanks for your interest in our work!

As described in Section 7.3 of our paper, the PanoWan dataset comprises not only the panoramic videos we collected from YouTube, but also several existing panoramic video datasets, including 360-1M, Imagine360, WEB360, 360+X, Panonut360, and the Miraikan 360-degree Video Dataset.

However, during data processing, the clip timestamps for the portion sourced from these existing datasets were not fully preserved. As a result, we have currently only released the records with available clip timestamps on HuggingFace, which include all the data we collected from YouTube.

That said, all the aforementioned existing datasets are publicly accessible on the internet. You can download them and apply the clipping, filtering, and cleaning procedures described in our paper to reproduce the remaining portion of the dataset.

Hope this helps!

[batman007herewego]

Thank you for your explanation. So, is Panowan's model fine-tuned and trained on Panovid based on a filtered 13K subset?