I was curious and clicked on the [Dataset] link and got this:<p><a href="https://github.com/transpchan/Live3D/issues/1" rel="nofollow">https://github.com/transpchan/Live3D/issues/1</a><p>and here's the paper review notes:<p><a href="https://github.com/transpchan/Live3D/blob/main/eccv.pdf" rel="nofollow">https://github.com/transpchan/Live3D/blob/main/eccv.pdf</a><p>which complains that it's too similar to<p><a href="https://anr-avatars.github.io/" rel="nofollow">https://anr-avatars.github.io/</a><p>and<p><a href="https://ai.facebook.com/research/publications/continuous-surface-embeddings/" rel="nofollow">https://ai.facebook.com/research/publications/continuous-sur...</a><p>To me, this looks like great craftsmanship: They have built a working implementation where others failed. And they found a way to cheaply generate lots of training data. But sadly, research optimizes for novelty, not for usability.