Research map / Vision-language and multimodal / Human and embodied modeling

Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization

ICML 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Composed and text-based image retrieval · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a query-optimization hashing method that uses learnable attribute queries with cross-attention decoding to produce attribute-aware hash bits, plus a training-only auxiliary branch (query circular-shift) to ease low-bit optimization for fine-grained image retrieval.

Problem

Learning-to-hash for large-scale fine-grained image retrieval performs poorly with low-bit hash codes, and hash bits lack interpretable attribute-level meaning.

Approach

Model hashing as set prediction: randomly initialized learnable queries decode attribute-specific features via cross-attention, each compressed into one hash bit; a training-only auxiliary branch reuses the decoder on circularly shifted queries to raise effective dimensionality without extra parameters, motivated by a cosine-similarity/Welch-bound analysis of large class counts in low dimensions.

Key results

• On CUB200 at 12 bits the method reaches 72.19% mAP vs 56.42% for AGMH, the strongest baseline in the table.
• Ablation on CUB200 at 12 bits: backbone-only 20.03%, +feature extractor 27.43%, +query learning 33.33%, +auxiliary branch 72.19%.
• Method has 31.71M params and 5.27G training FLOPs, lower than ExchNet, SEMICON, and AGMH in the efficiency table.

Measured effects reported in the paper

Surprising findings

• The auxiliary branch, which adds no inference parameters, raises CUB200 12-bit mAP from 33.33% to 72.19%, a much larger gain than the query-learning module itself.
• Without the auxiliary branch, the query-learning method is worse than prior state-of-the-art at low bit lengths, which the authors attribute to optimization difficulty rather than limited bit capacity.

Limitations

• Performance at 12 bits relies on a training-only auxiliary branch; the theoretical explanation is a lower-bound argument on mu rather than a full convergence analysis.
• Evaluated only on image retrieval with pairwise supervision and ResNet-50; cross-modal extension is left to future work.

Experimental setup

Datasets: CUB200-2011 FGVC Aircraft Food101 NABirds VegFru

Benchmarks: CUB200-2011 FGVC Aircraft Food101 NABirds VegFru

Metrics: mAP

Baselines: DPSH HashNet ADSH ExchNet A2-Net SEMICON A2-Net++ AGMH FISH CNET DAHN CMBH

Builds on: DPSH ExchNet A2-Net DETR AGMH

Closest papers by meaning

• A$^2$-Net: Learning Attribute-Aware Hash Codes for Large-Scale Fine-Grained Image Retrieval (NeurIPS 2021 · similarity 0.95)
• Learning Multifunctional Binary Codes for Both Category and Attribute Oriented Retrieval Tasks (CVPR 2017 · similarity 0.91)
• Characteristics Matching Based Hash Codes Generation for Efficient Fine-grained Image Retrieval (CVPR 2024 · similarity 0.90)
• An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing (CVPR 2024 · similarity 0.89)
• Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.87)
• Prototypical Hash Encoding for On-the-Fly Fine-Grained Category Discovery (NeurIPS 2024 · similarity 0.86)