Overview

This project predicts fashion trends for the Paris Fall/Winter 2025-2026 season by analyzing runway images from the Paris Fashion Week dataset.

It combines Detectron2 (ResNet-101 backbone) for clothing segmentation with FashionCLIP for fashion-specific attribute analysis. The result is a structured pipeline that extracts key fashion information from each image:

• Clothing items
• Dominant colors
• Visual attributes such as silhouette, pattern, and material

All results are saved in structured JSON format for downstream trend analysis and visualization.

check out the demo: notebooks/Demo.ipynb

Features

• Clothing Segmentation
  Uses a fine-tuned ResNet-101 Mask R-CNN model (Detectron2) to detect and segment garments like dresses, coats, tops, pants, etc.

• Dominant Color Extraction
  Applies KMeans clustering + perceptual filtering to extract the most representative colors from each segmented item.

• Attribute Analysis with FashionCLIP
  FashionCLIP is used to classify:

  • Silhouettes (e.g., A-line, bodycon, oversized)
  • Patterns (e.g., striped, floral, polka-dot)
  • Materials (e.g., leather, silk, fur)
    FashionCLIP was chosen over regular CLIP for its domain-specific training on fashion datasets.

• Output
  Each image is processed into a rich JSON file containing:

  • Metadata (designer, season, city, look)
  • Global attributes (silhouette, details)
  • Per-item analysis (type, pattern, material, color)