StreetStyle's Two-Part Engine: How One Framework Tracks Known Fashion Trends and Discovers New Ones

For street style photographers and trend analysts, spotting an emerging trend has traditionally meant standing at the right intersection with the right eye. A 2017 research project from Cornell University—StreetStyle, by Kevin Matzen, Kavita Bala and Noah Snavely—proposed a computational alternative: mine millions of street photos from social media, then let two distinct machine learning processes work in tandem. One tracks fashion attributes you already know to look for; the other surfaces coherent style groupings you never thought to define. Understanding that division of labor is the most useful idea fashion professionals can borrow from the project.

Introduction to the StreetStyle Framework

The project's premise is scale. Each day, the researchers observed, billions of photographs are uploaded to photo-sharing services and social platforms—images packed with information about how people live and dress. Rather than sampling from fashion week coverage, Matzen, Bala and Snavely exploited this trove of social media imagery to understand fashion and style trends worldwide, analyzing millions of photos of people spanning several years and many cities.

Clothing, as they note, is a demanding subject for computer vision: analyzing it requires both detailed attribute-level understanding and social context. The framework's response was a large-scale dataset of photos annotated with clothing attributes—labels such as "wearing glasses" or "short sleeves"—built on earlier vision research that describes people in terms of clothing attributes and recognizes those attributes in new images.

Part 1: Tracking Known Attributes with Supervised Learning

The first half of the engine is supervised learning. Researchers annotate a training dataset of photos with known clothing attributes, then use deep learning to train classifiers that predict those same attributes in new, unlabeled images. The process is "supervised" because a human decision precedes it: someone decides which attributes matter—black clothing, sleeve length, headwear—and the system learns to detect them consistently across millions of photos.

Crucially, this half answers predetermined questions. If you want to know whether short sleeves are rising in Seoul or whether black dominates in Los Angeles, you define the attribute first and let the classifier measure it at scale. The approach inherits the structure of the dataset's annotations: it can only track what it was explicitly trained to recognize.

Part 2: Discovering Novel Styles with Unsupervised Clustering

The second half addresses the question supervised learning cannot: what styles exist that no one listed in advance? The researchers present a method for discovering "visually consistent style clusters"—combinations of attributes and visual features that recur together and correlate usefully—without prior definitions. In their words, the paper sets out to identify common, visually correlated combinations of basic attributes (such as a blue sweater with a jacket and wool hat), styles that appear more frequently in one city than another or during particular periods, and finer-grained visually coherent versions of fashion elements.

Because no annotation step dictates what the clustering finds, this process can surface styles the research team never anticipated. The project page emphasizes this point with a concrete example: the system discovered the gele, the distinctive Nigerian head-tie, as an emergent cluster in its latent style space, despite never having been trained to recognize that specific headwear.

The Two-Part StreetStyle Analytical Engine

The framework's logic becomes clear when its components are laid side by side. The table below deconstructs StreetStyle into its pipeline, showing how each stage feeds the next and what each process contributes to the final output.

The two processes are complementary rather than redundant. Supervised classifiers yield quantifiable, comparable trend lines; unsupervised clustering yields unexpected groupings that would otherwise go unnamed. Together they produced the project's analysis of global and per-city fashion choices and spatio-temporal trends.

Practical Insights: Quantifiable Trends

The immediate payoff of attribute tracking is the ability to plot and analyze trends over time. The paper offers a representative question: how is the proportion of people wearing black in Los Angeles changing over time? Because the classifier labels every photo in the sample, the answer becomes a time series—an actual curve for a specific city and attribute, derived from millions of images rather than one observer's impressions.

This is where the method most directly augments manual trend-spotting. A photographer's archive answers what was seen where the photographer stood; the StreetStyle approach answers the same question with the same type of evidence, but at city scale and across multiple years.

Practical Insights: Localized Style Discovery

The clustering side produces a different order of insight. The researchers visualize style clusters as histograms whose patterns carry meaning: vertical stripes indicate a cluster concentrated in specific cities throughout the year, while horizontal stripes correspond to specific months or seasons. The gele cluster—very distinctive of Lagos, Nigeria and seen during much of the year—exemplifies the vertical case. Another cluster, a combination of fashion elements, appeared primarily during the Northern hemisphere's winter.

For a trend analyst, the lesson is that meaningful styles can be both hyper-local and seasonal, and that a discovery process not bound by pre-set categories is what reveals them.

Augmenting Trend Spotting with Computational Insights

The transferable idea from StreetStyle is not the software, which this article does not cover, but the two-pronged method: systematically track attributes you already care about, and separately run an open-ended search for visually coherent clusters you did not define. Fashion professionals can enhance their own trend analysis by conceptually adopting that structure—pairing deliberate, category-led measurement with data-driven observation of emergent, localized combinations. A concrete way to know the approach is working is the identification of a novel, geographically specific style cluster that was not previously part of established trend categories—the same kind of finding the Cornell researchers demonstrated when their system surfaced the gele without ever being trained to look for it.

Sources

• StreetStyle: Exploring world-wide clothing styles from millions of photos — ARXIV
• StreetStyle: Exploring world-wide clothing styles from millions of photos
• StreetStyle: Exploring world-wide clothing styles from millions of photos • MediaScopium