Avaliação de uma metodologia para recomendação de vestuário com base em similaridade utilizando a arquitetura YOLO e extração de características com DINOv2

Abstract

A recommendation system aims to suggest new items to the user based on their preferences. This work proposes evaluating a clothing recommendation methodology based on visual similarity, where the user inputs an image of a clothing item of interest. The YOLOv11 object detection model identifies all clothing items present in the image, and the user selects the item of interest, while the DINOv2 model is used for extracting its visual features. Next, the method involves performing a search for similar items in the DeepFashion2 database, using the Euclidean distance between the generated features. The results show that YOLOv11 achieved good mAP50 values, especially in experiments with a larger volume of training data. However, challenges were encountered in distinguishing visually similar classes, suggesting possible adjustments to the database. The study concludes that the methodology is viable for application in image-based clothing recommendation systems.