Business Problem

Modamia faced two primary challenges:

• Customization Overload: While allowing customers to design their own outfits was a unique selling point, many found it overwhelming and struggled to visualize how different pieces would come together.
• Lack of Personalized Recommendations: Traditional recommendation systems failed to provide personalized suggestions that catered to individual style preferences and body types, reducing customer satisfaction.

About Modamia

Modamia is revolutionizing the fashion retail space with AI-driven personalized shopping experiences, allowing customers to design customized outfits tailored to their style preferences, body types, and occasion needs.

Solution

Modamia partnered with goML to build an AI-powered solution that enhances the shopping experience through:

Photo-Based Outfit Recommendations: Using CNN models, customers could upload their photos and receive personalized outfit recommendations based on specific prompts like "I want a golden gown for a party." The AI would then generate a visual representation of the outfit on the user’s photo, leveraging MongoDB for real-time data processing.

‍

Human Body Measurement Estimation:
Utilized Efficient Net with added MLP layers and modified the final layer with a linear function to predict 3 girth body measurements. This involved input of front image, side image and added height and gender of given input. Here are the links to resources:
Training data : https://registry.opendata.aws/bodym/

Dynamic Adjustments to Preferences: The solution also allowed users to refine their choices by inputting specific preferences (e.g., "sleeveless"). The AI, powered by an MLP model, dynamically adjusted the outfit to meet these specifications, providing high-accuracy visual updates.

Enhanced Personalization & Visual Realism: By utilizing advanced CNN models, the AI system delivered realistic visual representations of how different outfits would look on each customer's unique body type. The front-end, built on React.js, ensured seamless user interactions.

Architecture

• Data Ingestion:
  Sources:
  CSV Measurements
  Images (Training Data)
  Storage: Both CSV and Image data are stored in cloud storage (depicted as green buckets).
• Preprocessing:
  CSV and Image data from storage undergo preprocessing, managed by a Lambda function (AWS Lambda).
  Human Body Measurement Estimation: Images are resized to (224,224). 3 channels are added as training images are grayscale and efficient Net requires 3 channels. Gender is one hot encoded and other inputs and outputs are normalized.
• Training Pipeline:
  The pre-processed data is passed to the training pipeline.
  The training process outputs a trained model and stores the model output in cloud storage.
• Testing Pipeline:
  The trained model is evaluated using test data.
  Output from the testing pipeline is also stored in cloud storage.
• Model Inference:
  Test Data is fed into an Inference API, which runs the preprocessing and testing processes.
  Final model output is generated and stored.

Training Pipeline: