AI-Powered Body Type Classification for Personalized Fashion Recommendations

AI-Powered Body Type Classification for Personalized Fashion Recommendations

Vaishali

Dept. of Computer Applications St Joseph Engineering College Vamanjoor, Mangaluru, India

Ms. Priyadarshini P

Assistant Professor

Dept. of Computer Applications St Joseph Engineering College Vamanjoor , Mangaluru, India

Mr. Hareesh B

Associate Professor Department of Computer Applications

St Joseph Engineering College Vamanjoor, Mangalore, Karnataka

Abstract – In the quickly evolving fashion technology sector of today's world, artificial intelligence (AI) and computer vision increasing play a important role in terms of enhancing user experiences and providing customized services. One significant aspect of this technological revolution is the automatic assessment of human body types to assist individuals in making better-informed fashion choices. Conventional body type evaluations frequently call for subjective assessments or manual measurements, which can be unreliable and inconvenient for the user. This study suggests a machine learning based body type classification system to simplify and customize fashion recommendations in order to address these issues.The system removes the need for human intervention by using sophisticated pose estimation techniques to extract precise body landmarks from user uploaded images.

The Pose solution from MediaPipe is used to effectively and precisely identify important human body landmarks. After being extracted, these features are categorized into numerical vectors and then fed into a Random Forest classifier, a popular machine learning algorithm that has been trained on a carefully selected dataset of pose landmarks labeled with different human body types. The system can automatically classify body types thanks to the combination of computer vision and supervised learning, giving users immediate body structure insights.The main objective of the suggested system is to help users understand their body types and offer wardrobe recommendations that complement their physical characteristics, thereby increasing their overall fashion confidence.Experiments show that the model predicts body types with dependable accuracy, confirming its potential as a workable solution in the customized fashion industry. The study also demonstrates how scalable, approachable solutions that bridge the gap between technology and fashion personalization can be produced by combining lightweight computer vision pipelines with traditional machine learning algorithms. The proposed approach paves the way for future advancements in AI-driven fashion applications, virtual styling assistants, and intelligent wardrobe management.

Index Terms – Artificial Intelligence, Computer Vision, Body Type Detection, Fashion Personalization, Pose Estimation, Random Forest Classifier, Human Body Analysis

1. INTRODUCTION

   In the evolving fashion industry, personalized styling and outfit recommendations have become increasingly important for

   enhancing customer experiences and improving confidence in clothing choices. Traditionally, identifying ones body type has been a manual process, often relying on subjective opinions or physical measurements. However, these methods are time consuming, prone to error, and inaccessible to individuals without fashion expertise. With the growing influence of artificial intelligence (AI) and computer vision, there is an opportunity to automate and simplify this process, making personalized fashion advice more accessible to everyday users.

   A person's body type is one of the most important considerations when choosing appropriate attire. Different body types such as pear, hourglass, inverted triangle, and rectangle influence how clothes fit and flatter a person. When users are unaware of their body structure, they may struggle to choose outfits that complement their physique, leading to dissatisfaction with their wardrobe selections. Addressing this challenge, automated body type analysis can play a transformative role in guiding users toward better fashion decisions.

   This paper proposes an AI-enabled body type classification system that combines computer vision-based pose estimation with a machine learning classifier to accurately identify body types from user images. The system uses the MediaPipe Pose solution to extract key body landmarks from an image and processes these landmarks to form a feature vector. This data is then fed into a Random Forest classification model, which predicts the user's body type based on learned patterns from a curated dataset. By automating the body type detection process, the system assists users in understanding their physical structure without manual effort and delivers personalized outfit suggestions aligned with their body shape.

   The proposed approach bridges the gap between technology and fashion by using lightweight, accessible AI tools that do not require expensive hardware or complex measurements. The goal is to empower individuals with better wardrobe choices while

   simplifying the fashion selection process using user friendly technology. This research contributes to the growing field of AI- driven fashion assistants and lays the foundation for future enhancements in virtual styling platforms and smart wardrobe applications.

2. LITERATURE REVIEW

   The fashion tech intersection has seen remarkable growth thanks to advances in computer vision and machine learning. This review highlights recent works in pose estimation, body shape analysis, and AIdriven fashion personalization, and frames how your research fills a gap.

   1. Advanced Pose Estimation for Human Shape

      Pose estimation remains a critical foundation for body analysis. A recent comprehensive review by Almquisadi et al. (2023) surveys deep learning models for 3D human pose estimation, covering both CNN and transformer based approaches while discussing real world challenges like occlusion and crowded scenes .These system,designed for robust landmark detection provide the necessary structure to extract meaningful body features.

   2. Body Shape Recovery and Classification

      Theres growing interest in estimating body shape from RGB images. ShapeBoost (Bian et al., 2024) introduces a part based parameterization framework along with clothing preserving data augmentation to recover body shape even under varied attire significantly improving accuracy on diverse shapes. Complementing this, Shape of You (Sarkar et al., 2023) propels 3D body reconstruction by formulating new loss functions and testtime optimization routines, achieving a 17.7% performance gain over prior methods on the SSP3D dataset .

      Even more relevant is Trotter et al. (2023), who presented a system capable of classifying human body shape categories (e.g. rectangle, pear, hourglass) from a single image, leveraging instance segmentation and keypoint estimation no 3D modeling required. Their work is among the first to focus on direct shape classification from pose landmarks.

   3. Fashion Personalization and Virtual Try-On Systems

      AIpowered fashion applications are incorporating body aware mechanisms more often. While systems like FashionFit (Mohammadi & Kalhor, 2022) integrate pose estimation and segmentation to generate realistic virtual tryon experiences, they still rely predominantly on garment retrieval and style matching rather than explicit body shape inference .

      Comprehensive surveys of fashion AI show that virtual styling platforms largely omit body-type detection as an input favoring user behavior and apparel metadata instead. This reveals an opportunity to bring body structure into the personalization loop.

   4. Research Gap

      Desite substantial progress in pose estimation and 3D shape reconstruction, the fast-growing field lacks direct application of poselandmark based classification of body types. Virtual tryon and recommendation systems typically dont integrate automatic body type detection to inform styling suggestions a significant gap this work addresses.

   5. Your Contribution

   This papers approach leverages lightweight yet powerful tools like MediaPipe Pose for real-time landmark extraction and a Random Forest classifier trained on curated poselandmark data to automatically identify body type categories. This complements recent 3D shape estimation systems by offering a practical and accessible solution, directly supporting personalized wardrobe advice based on body morphology.

3. METHODOLOGY

   1. Model Architecture and Functions

      The proposed system is built on a two stage architecture combining pose landmark detection and machine learning based body type classification. Together, these two elements work together to automatically analyze a persons body shape from an image and predict the corresponding body type.

      The first stage involves MediaPipe Pose, an efficient and lightweight pose estimation model developed by Google. This model takes a user uploaded image as input and accurately identifies 33 key human body landmarks, including the shoulders, hips, elbows, and knees. These details are essential for understanding the users body proportions. MediaPipe Pose was selected because of its real time performance and its ability to handle varying lighting conditions, orientations, and backgrounds, making it ideal for user submitted images.In the second stage, the extracted pose landmarks are transformed into a numerical feature vector. This vector is fed into a Random Forest Classifier, a supervised machine learning model known for its reliability on structured data and its ability to handle complex, non-linear relationships. The Random Forest model was trained to classify users into one of the defined body types, such as Pear, Rectangle, Hourglass, or Inverted Triangle, based on their landmark patterns.

      Both models are connected through a Flask based API, which handles image processing, landmark extraction, body type prediction, and sends the results back to the frontend in real-time. The system is designed to be lightweight and scalable, allowing seamless integration into the personalized fashion recommendation application.

      Fig. 1. Overall architecture of the proposed body type classification system integrating MediaPipe Pose and Random Forest Classifier.

   2. Dataset Preparation and Processing

      To train the body type classification model, two key steps were involved: data collection and preprocessing.

      • Data Collection: A dataset was created using human images with varying body shapes. MediaPipe Pose was used to extract landmarks from these images, resulting in a structured dataset where each sample represented a person's body through landmark coordinates (x, y, z) and visibility scores. Each entry is labeled with its corresponding body type, forming the ground truth for supervised learning.

      • Preprocessing:

        • Landmark coordinates were normalized to minimize the effect of image resolution differences.

        • Missing or low-confidence landmark detections were addressed using thresholding based on the visibility score.

        • The data was flattened into one-dimensional feature vectors, where each element represented a specific coordinate of a body landmark.

        • The dataset was split into training and test sets (80% and 20%, respectively) to evaluate the model's performance.

        • Data shuffling was performed to ensure balanced training.

      This structured and labeled dataset provided the foundation for training the Random Forest classifier to recognize patterns unique

      to each body type.

      Fig. 2. Flowchart of data preprocessing and feature extraction

      for the Random Forest model.

   3. Evaluation Metrics

      The performance of the proposed system was evaluated using the following metrics:

      • Body Type Classification Accuracy: Measured as the percentage of correctly predicted body type labels on the test dataset. This gives a direct indication of how well the model distinguishes between different body types.

      • Pose Detection Robustness: While MediaPipe Pose itself is pre-trained, qualitative checks ensured that the model consistently extracted all 33 landmarks across diverse images without missing critical points.

      • User Testing Feedback: Initial feedback was collected from users who tested the system with their images and verified whether the predicted body type aligned with their actual body shape. This subjective evaluation gave insights into the models practical usability.

4. CONCLUSION

   This study presents a practical approach to automating body type classification using pose landmark detection and machine learning. By combining MediaPipe Pose for efficient human pose estimation and a Random Forest classifier for body type prediction, the system successfully addresses the challenge of providing users with personalized fashion insights without the need for manual measurements or subjective assessments.

   The proposed solution is lightweight, accessible, and designed to integrate seamlessly into personalized fashion applications. Through automated landmark extraction and structured classification, the system empowers users to understand their body type and make more informed wardrobe choices. Experimental results indicate that the system can predict body types with promising accuracy, making it a useful tool for enhancing wardrobe management and improving virtual styling recommendations.

   In a broader sense, this work demonstrates how advancements in computer vision and machine learning can be applied to everyday problems, making fashion personalization more intelligent and user centric. Future enhancements to the system could include expanding the dataset for better generalization, exploring deep learning classifiers for improved accuracy, and integrating real- time outfit recommendations based on detected body types.

   Overall, this research contributes to the growing field of AI- enabled fashion technologies, offering a foundation for more interactive and tailored fashion experiences.

5. FUTURE SCOPE

   While the proposed system demonstrates the ability to classify body types accurately using pose landmarks and machine learning, there are several opportunities to expand and enhance this work in the future.

   One important direction is to improve the accuracy and robustness of the classification model by training on a more diverse and larger dataset. Including images of people with varied clothing, lighting conditions, body postures, and demographics can help the model generalize better to real-world scenarios.

   Another potential enhancement is the integration of deep learning-based classifiers, such as Convolutional Neural Networks (CNNs) or Transformer models, which could further improve classification accuracy. These models might better capture subtle variations in body structure beyond what is possible with traditional machine learning techniques.

   In the context of fashion personalization, the system can be

   extended to provide real-time outfit recommendations tailored to the detected body type. By analyzing clothing databases and applying style guidelines suited to each body type, the system could automatically suggest outfits that best flatter the users physique.

   Future work could also invole building a mobile application for ease of access, allowing users to upload images directly from their smartphones and instantly receive body type analysis and fashion suggestions. Additionally, adding modules like virtual try-on and wardrobe planning would create a complete personal fashion assistant.

   Lastly, integrating user feedback mechanisms will help continuously improve the system's predictions and outfit suggestions by learning from actual user preferences and satisfaction levels.

   Overall, these future enhancements aim to make the system more personalized, adaptive, and aligned with modern digital fashion trends.

6. REFERENCES

1. R. Almquisadi, A. Alzubi, A. Sharieh, A. Tubaishat, and A. Abu-Shareha, A systematic review of recent deep learning approaches for 3D human pose estimation, Artificial Intelligence Review, vol. 57, no. 2, pp. 12351260, Feb. 2024.

2. S. Bian, X. Xu, X. Li, and Z. Lin, Shapeboost: Boosting human shape estimation with part-based parameterization and clothing-preserving augmentation, arXiv preprint, arXiv:2403.01345, Mar. 2024.

3. R. Sarkar, A. Sanyal, and S. Sengupta, Shape of you: Precise 3D shape estimations for diverse body types, arXiv preprint, arXiv:2304.07389, Apr. 2023.

4. C. Trotter, Y. Li, and A. Jain, Human body shape classification based on a single image, arXiv preprint, arXiv:2305.18480, May 2023.

5. S. Mohammadi and A. Kalhor, FashionFit: Analysis of mapping 3D pose and neural body fit for custom virtual try-on, Procedia Computer Science, vol. 188, pp. 121128, 2022.