Image Retrieval through Retrieval-Oriented Dimensionality Reduction

Large-scale fashion image retrieval relies on high-dimensional embeddings that improve accuracy but introduce significant computational and storage costs. Dimensionality reduction offers a solution, yet many methods degrade retrieval quality by failing to preserve neighborhood structure. We evaluate linear, nonlinear, and neural approaches across state-of-the-art backbones on the DeepFashion In-Shop dataset, finding PCA to be a strong baseline while naive autoencoders perform poorly. To address this, we propose the MCDO Loss, a hybrid objective combining reconstruction fidelity, cosine similarity preservation, bottleneck decorrelation, and orthogonality regularization. Experiments show that AE-MCDO consistently outperforms PCA and MSE-based autoencoders, achieving higher Hit@K with more compact representations, thus enabling efficient and accurate large-scale fashion search.