Decoding the fashion trend of sports shoes with empowered computer vision

Qian Lu, Jingjing Li, Ziseng Lin, Jin ZhouIndustria 2023 · DOI

TL;DR

This paper presents an automated pipeline to replace subjective expert intuition in footwear forecasting with objective computer vision. By combining Mask R-CNN for style classification and a perceptually-aware K-means clustering for color analysis, the researchers successfully decoded 2021 fashion trends from over 11,000 e-commerce images, identifying "Atlantic Blue" and "Red Ochre" as key seasonal markers.

What Problem Does This Paper Address?

The fashion industry is notoriously volatile. With the rise of "fast fashion," product life cycles have shrunk, and consumer tastes shift faster than traditional supply chains can react. Historically, brands have relied on "fashion forecasters"—experts who use intuition to predict next season's hits. However, this process is slow, expensive, and prone to human bias.

While computer vision (CV) has been widely applied to clothing (shirts, dresses, etc.), footwear presents unique challenges. Shoes are smaller, have more complex geometric silhouettes, and often feature intricate color-blocking that is harder to categorize than a solid-colored garment. This paper seeks to bridge that gap by creating a systematic, AI-driven way to analyze footwear trends at scale.

Methods

The researchers developed a two-stage computer vision system:

  1. Style Classification and Segmentation: The team used Mask R-CNN (with a ResNet101-FPN backbone). Unlike standard object detectors that just draw a box around an object, Mask R-CNN performs "instance segmentation," creating a pixel-perfect mask of the shoe. This allows the system to ignore the background (like shadows or floor textures) and focus purely on the shoe's attributes.
  2. Color Extraction: Instead of using standard RGB values—which don't align well with how humans perceive color—the researchers converted images to the CIELAB color space. CIELAB is "perceptually uniform," meaning the numerical distance between two colors matches the perceived difference to the human eye.
  3. Clustering: They applied a modified Mean-Ranked K-means clustering algorithm to the extracted shoe pixels to identify the dominant "main" colors of each model, eventually grouping them into 25 distinct color clusters.

Datasets

The authors built a custom dataset for this study:

  • Total Images: 11,267 sneaker images scraped from major e-commerce platforms (SSENSE, Farfetch, Nike, Adidas, etc.) covering the full year of 2021.
  • Annotations: 4,744 images were manually annotated with precise masks using LabelMe to train the segmentation model.
  • Categories: The data was categorized into five styles: Fashion sports shoes, Casual sports shoes, Board shoes, Training shoes, and Basketball shoes.

Key Insights and Ideas

  • The Dominance of the "Low-Top": Across almost every category (except basketball), low-top silhouettes represented the vast majority of the market. For fashion sneakers, low-tops accounted for over 81% of the data, suggesting that consumer preference for versatility and ease of wear outweighs "avant-garde" high-top designs.
  • Seasonal Color Logic: The AI confirmed that footwear follows the same "seasonal laws" as apparel. Winter shoes trended toward dark greys and blacks (46% of the palette), while spring and summer saw a shift toward "cool" light greys and vibrant tones.
  • Validation against Industry Giants: The system identified "Atlantic Blue" and "Reddish Orange" as trending colors for 2021. These findings directly aligned with professional trend reports from WGSN (a leading global trend forecasting agency), proving that the AI could replicate expert insights autonomously.
  • Neutral vs. Vibrant: While neutral tones (grey, white, black) dominate the volume of sales, the "fashion" signal is found in the 0.5%–5% of vibrant colors like red ochre, which act as the actual trend indicators.

Key Results

  • Model Performance: The Mask R-CNN achieved an average similarity (Intersection over Union) of over 85.5% for shoe recognition.
  • Efficiency: The inference time was approximately 0.1 seconds per image, making it orders of magnitude faster than manual human auditing.
  • Market Breakdown: "Fashion sports shoes" (40.12%) and "Casual sports shoes" (24%) were the two most prevalent styles in the 2021 market.
  • Color Accuracy: The CIELAB-based clustering successfully mapped 25 distinct color trends, identifying specific popular hues like "Atlantic Blue" in board shoes and "Red Ochre" in basketball shoes.

Limitations and Future Work

While the study is a strong proof-of-concept, the authors note several areas for growth:

  • Data Diversity: The current dataset relies on e-commerce images, which are often "clean" and staged. Future work should include "street style" photos from social media and runway images to capture trends as they emerge in the wild.
  • Attribute Depth: The current model focuses on style and color. The authors suggest that future iterations should include material recognition (leather vs. suede vs. mesh) and pattern analysis (logos, prints, textures).
  • Trend Propagation: The researchers aim to study how a trend moves geographically and temporally—for example, how a color trend moves from a high-fashion runway to a mass-market e-commerce site.

Summary updated

Abstract

To adapt to the rapidly changing market and capricious trends, fashion brands need to understand trends and market conditions precisely and fast to produce marketable products. The traditional fashion trend analysis has relied heavily on the subjective judgement of experts, inevitably leading to biased decisions and is time-consuming. The development of computer vision and machine learning provides an objective and systematic approach to processing images and analysis of fashion products. However, most studies focus on clothing trends analysis, few are on shoe trends analysis. Hence, this study aimed to decode and analyse the fashion trend of sports shoes with empowered computer vision technology. In this paper, a dataset containing e-commerce images of sports shoes with precise annotations was established; then Mask-RCNN was utilized to classify and extract the shoe from the background image; a modified version of the K-means clustering algorithm was employed to detect the shoe colour. The results indicated that fashionable sports shoes and casual sports shoes were the most prevalent two styles. Besides neutral tones, yellow, red ochre and reddish orange were popular in casual sports shoes, fashion sports shoes and basketball shoes respectively and Atlantic Blue in board shoes and trainers. This study demonstrated the promising potential of computer vision and machine learning as a new method to analyse footwear fashion trends efficiently and economically.
BibTeX
@inproceedings{lu2023decoding,
  author = {Qian Lu and Jingjing Li and ZISENG LIN and Jin Zhou},
  title = {{Decoding the fashion trend of sports shoes with empowered computer vision}},
  booktitle = {Industria},
  year = {2023},
  doi = {10.35530/it.074.05.2022107},
}
Figure