Benchmarking Dimensionality Reduction Techniques for ViT Features

Accuracy, Storage, and Latency Trade-offs in Edge AI Image Retrieval

Abstract

Content-Based Image Retrieval (CBIR) systems face a major challenge today in the era of Edge AI. Vision Transformers (ViT) generate high-dimensional feature vectors (768 dimensions). This creates serious memory and pro-cessing problems for devices with limited resources. This empirical study tests three different ways to reduce these dimensions: Principal Component Analysis (PCA), Autoencoders (AE), and Product Quantization (PQ). These methods were applied to ViT features using a merged dataset of Oxford5k and Paris6k, containing 11,475 images. The study focuses heavily on real-world constraints like storage size, search latency, and accuracy.
To ensure the findings are statistically robust, all experiments were repeated 10 times. The results show a clear trade-off. Traditional linear methods like PCA lose too much accuracy under heavy compression, dropping to 0.4907 mAP at 64 dimensions. On the other hand, Autoencoders keep the semantic quality much better, reaching a mean of 0.5106 mAP. Meanwhile, Product Quantization proved to be highly storage-efficient. It achieved a 47.9x com-pression ratio while maintaining a competitive accuracy of 0.5087 mAP.
Based on these results, the paper suggests PQ-128 as a strong candidate for systems with limited storage. It provides a 24x compression ratio with a very small 0.56% drop in accuracy. The paper aims to help engineers choose the best compression method based on their specific device constraints. This guide is useful for deploying high-demand local applications, such as smart surveillance systems and high-speed visual search, on affordable hardware.