Ethnic Fit: Siamese Network-Based Outfit Compatibility Prediction for Indian Ethnic Wear

Abstract

Outfit compatibility prediction has been studied extensively for Western clothing using the Maryland Polyvore benchmark dataset, with state-of-the-art models such as OutfitTransformer achieving AUC scores of 0.92 (Sarkar et al., 2023). However, no published work addresses this problem in the context of Indian ethnic wear, which includes sarees, kurtas, lehengas, sherwanis, and salwar suits. These garments differ significantly from Western clothing in terms of draping conventions, layering patterns, and occasion-based colour pairing norms. This paper presents Ethnic Fit, a Siamese Network-based deep learning model for pairwise outfit compatibility prediction in Indian ethnic wear. A pretrained ResNet-50 backbone is used to extract 512-dimensional visual embeddings from garment images, and the absolute difference between embeddings is passed through fully connected classification layers trained with Binary Cross Entropy loss and the Adam optimiser. The model is trained on a combined dataset comprising 9,434 Maryland Polyvore pairs and 600 newly curated Indian ethnic garment pairs drawn from the Indo Fashion Dataset, covering 15 categories. Evaluated on a held-out Indian ethnic wear test set, the proposed model achieves AUC 0.9553, Accuracy 91.67%, and F1-Score 0.9000, establishing the first published benchmark for this domain. A live Gradio-based inference interface is also deployed for real-world testing.