AI Customization for Comprehensive Fashion Classification

Looking for DL/CV engineers with relevant image/video classification AI model customization experience. Preferably in fashion-related projects. We are a pre-seed fashion tech start-up aiming at creating innovative tools to assist users' styling decision-making process and contribute values to fashion sustainability. We aim to establish long-term working relationships with engineers that help us on this project. Technical Expertise: • Machine Learning/Deep Learning: image classification and analysis (e.g., CNNs, ResNet, InceptionV3, EfficientNet). • Computer Vision: feature extraction, object detection, and image processing. • Data Science: ability to work with large datasets, including data cleaning, augmentation, and analysis. Knowledge of statistical analysis and predictive modeling. Tools and Technologies: • Python, TensorFlow, PyTorch, Keras, and OpenCV. • Dataset Handling. Experience in managing and processing large and diverse datasets • AI Model Optimization, such as model pruning, quantization, and distillation. • Data protection knowledge such as GDPR and know how to implement privacy-by-design principles. Project Overview Objectives • Customization and Enhancement: Adapt a pre-trained AI model to accurately identify and classify specified fashion design details in images or videos. • Innovation in Category Prediction: Develop a mechanism for the AI to autonomously predict and create new fashion categories that are not pre-defined in our existing dataset. Technical Requirements I. Existing AI Model Choices and Customization Approach Objective: The primary goal is to customize a pre-trained AI model to accurately identify, classify, and tag detailed fashion design elements in images or videos. These elements include but are not limited to collar types, shirt colors, and other intricate design details. The model should also be capable of handling images of varying resolution, varying lighting settings, including less ideal ones, with high accuracy and detail detection. Furthermore, it should support tagging third-party online retailer product images and matching these with our datasets. Initially, the model will be used offline for research, analysis, and tagging of new arrivals daily. However, it is intended to eventually scale the model for real-time application. Pre-trained Model Options: (Open to proposals) • ResNet • InceptionV3 • EfficientNet • Mask R-CNN • DenseNet II. Autonomous Category Prediction and Creation Objective: Implement an advanced feature within the AI model that enables the autonomous prediction and creation of new sub-categories for a broad range of fashion items, including clothing, shoes, and accessories, when such items do not align with existing classifications in the dataset. Requirements: • Innovative Detection Mechanism: Implement an advanced detection mechanism within the AI model that can identify when an input (image/video) contains fashion details that do not fit into pre-defined categories in our dataset. • Predictive Modeling for Category Suggestion Objective: Equip the AI model with the capability to identify items across all categories—clothing, shoes, and accessories—that do not fit into predefined sub- categories. The model should suggest new, relevant sub-categories based on an in- depth analysis of features and patterns that differentiate these items from existing classifications. - Comprehensive Feature Extraction: Implement state-of-the-art convolutional neural networks (CNNs) or equivalent architectures to perform deep feature extraction. This process should capture the diverse aspects of fashion items, such as design, patterns, color, and other relevant attributes to accurately categorize items. - Anomaly Detection and Suggestion Mechanism: Use anomaly detection to highlight items that fall outside known sub-categories. Based on the distinctive features of these items, the AI will propose new sub-categories, enriching the existing dataset with nuanced classifications that better reflect the variety of fashion items encountered. - Contextual Descriptions for New Categories: For each suggested new sub- category, the model should provide detailed descriptions highlighting the unique attributes and reasons behind the suggestion. • Expert Review Mechanism for Validation and Feedback - Interactive Validation Interface: Create an intuitive interface for fashion experts to evaluate the AI's suggested new sub-categories across all fashion items. - Incorporating Expert Feedback: Establish a feedback loop where experts can approve, modify, or reject the suggested new sub-categories. - Documentation and Continuous Improvement: Document all expert-validated new sub-categories, including their defining characteristics and exemplary items. Regularly update the dataset and the AI model with this new information to continuously enhance the model's predictive accuracy and its alignment with current fashion trends.