Indian Fashion Insights framework

Pioneering deep learning in fashion with a custom Indian dataset for advanced trend forecasting and targeted consumer insights.

Fashion Retail

2022

Client and Industry

The client is a leading institution in India renowned for its comprehensive fashion education, offering courses in design, management, and technology. Its influential role significantly shapes the fashion industry both domestically and internationally.

Industry

Fashion Retail

Business Type

Institute

Services

Research & Development

Our Role

We developed a novel deep learning-based approach for attribute-level multi-label classification of a specially curated Indian fashion dataset. This initiative is designed to predict fashion trends and consumer behaviour in the Indian market.

Project Goals

01

Dataset Creation: Construct a bespoke dataset from scratch, sourcing images from social media, e-commerce sites, and various online resources.

02

Data Annotation: Meticulously annotate the dataset to capture a wide range of attributes including design elements, fabric types, colors, and patterns.

03

Model Development: Engineer a specialized two-stage deep learning model that integrates a detection component with a classifier model ensemble for enhanced accuracy.

04

Insight Generation: Produce authentic, geographically and temporally specific fashion insights to empower the Indian textile, fashion, and retail industries in developing products tailored to the Indian market.

Business
Challenges

01.

Leveraging a dataset that encompasses approximately 125 categories, including 15 types of accessories, using around 600,000 unique images for model training.

02.

Providing annotations for each image at both the category and attribute levels, with special attention to unique cases such as sarees and one-piece garments.

03.

Implementing robust detection capabilities for multiple accessories across various scenarios was also critical.

Solution

We developed a highly accurate detection model tailored for traditional Indian ethnic garments, achieving 99% accuracy in identifying sarees and one-piece outfits. This model is part of a broader high precision, low latency prediction platform that efficiently processes and interprets fashion data.

Key Features

01

The system incorporates a sophisticated multihead deep learning architecture that supports concurrent training on multiple clothing attributes. This architecture uses a custom loss function to enhance model accuracy, effectively reducing the need for multiple models to just one.

02

A finely-tuned YoloV7 model within the data pre-processing pipeline ensures accurate detection of individuals in various attire types. The model excels in fine-grained attribute classification by incorporating localized attention through on-body key points, resulting in heightened precision in identifying clothing details.

03

For accessories, we implemented a unified accessory flagger. This single model adeptly identifies all accessories present in an image, streamlining accessory detection across various scenarios.

04

The platform also compiles comprehensive insights from all processed and inferred data, supporting the creation of mood boards and storyboards for detailed fashion analysis.

Impact

01.

Enhanced Forecasting and Responsiveness: The insights allowed our client to develop a comprehensive app tailored to the current technological framework, enhancing their ability to predict and adapt to evolving fashion trends effectively.

02.

Industry Leadership and Value Addition: This development bolstered their position as a thought leader in the fashion education space and provided invaluable tools for industry professionals aiming to stay aligned with the rapid changes in the Indian fashion landscape.