Optimizing Approximate Nearest Neighbor Search for Fashion Image Retrieval

In this study, we conduct a comparative evaluation of the Hierarchical Navigable Small World (HNSW) algorithm, as well as Inverted File Index (IVF) and Product Quantization (PQ), analyzing their impact on search accuracy, indexing time, and query performance. We experiment with embeddings extracted from the Fashionpedia dataset using ResNet-101, with additional dimensionality reduction via Principal Component Analysis (PCA) for performance comparison. Furthermore, we incorporate category-based filtering to enhance retrieval relevance. Our findings provide insights into the trade-offs between recall, efficiency, and memory usage, offering practical guidelines for optimizing ANN-based indexing techniques in visual recommendation tasks.