The author reflects on the different approaches to animation quality at CVPR and SIGGRAPH, noting tensions between the standards of the vision and graphics communities. They emphasize the importance of high-quality animation data, comparing datasets to showcase the gold standard. They delve into the significance of temporal resolution and numerical precision in capturing realistic motion, highlighting errors that can arise from insufficient detail. The author explores issues with existing datasets, pointing out systematic errors, glitches, and limitations that can impact neural network training. They caution against mixing datasets of varying quality, arguing that “garbage in, garbage out” prevails in deep learning.
https://theorangeduck.com/page/animation-quality