"StreetStyle: Exploring world-wide clothing styles from millions of photos"

Hey everyone,

I just came across a paper that explores an interesting twist on fine-tuning a CNN - training a binary classifier for each of many “clothing styles” using some near-final layer CNN activations where the authors scraped millions of geo-tagged images from Instagram and spent $4k on mechanical turk to get clothing style annotations

The paper also does some interesting clustering across countries and is a good example of an end-to-end analysis (including some ways to make sure you gather a quality dataset using mturk)

4 Likes