This reminds me of a paper put out by a few researchers at Google Brain: `No Classification without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing World` [0]
The take away was that existing open image datasets are biased toward western contexts (eg what a wedding looks like), leading to low performance when applied in non-western contexts.
The take away was that existing open image datasets are biased toward western contexts (eg what a wedding looks like), leading to low performance when applied in non-western contexts.
[0] https://arxiv.org/abs/1711.08536