Reproduce “End-to-End Supervised Product Quantization for Image Search and Retrieval” on OpenResearch

About this paper

Benjamin Klein, Lior Wolf

Researchers at Tel Aviv University and Facebook AI Research developed Deep Product Quantization (DPQ), an end-to-end supervised method that learns dictionary-based compressed representations for image search. DPQ achieved state-of-the-art results on several benchmarks, including 0.7543 mAP for 24-bit retrieval on CIFAR-10 and 56.80% Top-1 accuracy for ImageNet classification with 64-bit representations.

Key insights

• Integrating supervised learning directly into the quantization process through end-to-end optimization significantly enhances the discriminative power of compressed representations.
• The Joint Central Loss effectively bridges the gap between continuous 'soft' and discrete 'hard' representations, improving the accuracy of asymmetric search by ensuring their consistency.
• Intra-normalization of sub-vectors and careful feature preprocessing (e.g., L2-normalization for PQ-Norm baseline) are crucial for robust performance, especially in cross-domain retrieval tasks.

Summary by alphaXiv.