GeoStyle: Discovering Fashion Trends and EventsUtkarsh Mall1 Kevin Matzen2 Bharath Hariharan1 Noah Snavely1 Kavita Bala1

Pipeline

Results

Future Work

Contributions

● Extending the method for other domains and datasets.● Analysis at a fine-grained level.● Looking at methods to alleviate bias due to dataset.

Detecting and predicting fashion trends over space and time● A fully-automated framework for fashion trend discovery.● Accurately model and forecast long-term trends, like seasonal

cycles.● Discover and group short-term spikes caused by events, like

rallies, festivals and sporting events● Automatically discover the reasons for these events.

Stage II - Characterizing Trends

Stage III - Discovering Events

Stage IV - Mining Reasons for Events

44 cities3 years (June 2013-May 2016)7.7 million imagesInstagram+Flickr100m

Bangkok Seattle Bangkok Chicago Rio de Janeiro

Yellow Color No Sleeves T-shirt Red Color Yellow Color

2014 Dec 2015 Dec

2014 Oct 2014 Apr 2014 Jun 2014 Jun2014 Jul

dadfather

halloweenfreaknight

songkranfestival

cupstanleycup

worldcupbrasil

Discovers events even in held-out cities from a different domain (Flickr).

● Classify attributes from StreetStyle [1].● Aggregate probability values at each time & city.

● Use TF-IDF on captions to identify words most representative of an event.

● Find outliers using hypothesis testing framework.● Group outliers based on proximity and periodicity.

4th, july, pride, america, 4thofjuly

halloween, freaknight, freaknight2014 happyhalloween, halloween2014

1Cornell University 2Facebook

Stage I - Recognition

AcknowledgmentThis work was funded by NSF (CHS: 1617861 and CHS:1513967) and an Amazon Research Award.

Accurate, fine-grained prediction of trends Discovered new events unknown to the authors

newyear, winter, christmas, holidays, year

Attribute Recognition&

Style Discovery

Take-away● Modelling trends using prior understanding results in more

accurate and interpretable forecasting models.● 20% improvement in prediction error rate v/s VAR.● A total of 725 events discovered of which 456 are interpretable

as seasonal sporting events, festivals or political rallies. ● Text combined with visual features helps in explaining events.

● 100% of the top 35 events are explainable by captions.

“Catalan way”2013 Sep in Barcelona

Lower is better

Popular events around the world using representative styles

● Interpretable, expressive parametric model capturing linear and cyclical trends.

(I)

(II)

(III)

(IV)

t-SNE of dataset images

References[1] K. Matzen, K. Bala, and N. Snavely. Streetstyle: Exploring world-wide clothing styles from millions of photos. CoRR, 2017

Parameter Interpretable meaning

r Trade-off between linear and cyclic trend

mlin Rate of long-term change in popularity

mcyc Amplitude and sign of cyclical spikes

k Spikiness of cyclical spikes

Styles are discovered by clustering in embedding space [1].