Research map / Label-efficient and robust learning / Self-supervised and contrastive
An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing
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 A2-SSL, an unsupervised fine-grained image hashing method combining asymmetric augmentation contrastive learning, part-oriented dense contrastive learning with Fisher Vectors, and a self-consistency reconstruction loss for hash codes.
Problem
Unsupervised hashing methods perform far worse on fine-grained image datasets than on generic ones (a "granularity gap"), because conventional contrastive SSL with strong augmentations can destroy fine-grained category cues.
Approach
Use asymmetric augmentation (mild views for anchor/positive, destructive ColorJitter+ElasticTransform for negatives), add part-level dense contrastive learning over GMM/Fisher Vector descriptors, and a self-consistent encoder-decoder reconstruction loss on the hash embedding.
Key results
- On CUB200-2011 at 48 bits, A2-SSL reaches 29.19% mAP versus 18.80% for MeCoQ (best baseline at that setting in Table 2).
- On Stanford Dogs at 48 bits, A2-SSL reaches 64.56% mAP versus 53.76% for MeCoQ.
- Ablation on CUB200-2011 (12-bit) shows asymmetric augmented SSL raises mAP from 8.41% to 17.62%, and full model reaches 20.14%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use A2-SSL instead of MeCoQ | Unsupervised fine-grained image hashing · CUB200-2011, 48-bit | mAP (%): 18.80 -> 29.19 | Helped |
| Use A2-SSL instead of MeCoQ | Unsupervised fine-grained image hashing · Oxford Flowers, 48-bit | mAP (%): 35.44 -> 46.37 | Helped |
| Use A2-SSL instead of SPQ | Unsupervised fine-grained image hashing · Stanford Dogs, 48-bit | mAP (%): 56.55 -> 64.56 | Helped |
| Use A2-SSL instead of SPQ | Unsupervised fine-grained image hashing · Stanford Cars, 48-bit | mAP (%): 4.36 -> 5.28 | Helped |
| Use A2-SSL instead of SPQ | Unsupervised fine-grained image hashing · Food101, 48-bit | mAP (%): 13.33 -> 15.38 | Helped |
| Add asymmetric augmented SSL instead of conventional contrastive learning (ablation) | Unsupervised hashing · CUB200-2011, 48-bit | mAP (%): 15.36 -> 26.57 | Helped |
| Add part-oriented DenseCL without cluster weights (ablation) | Asymmetric augmented SSL hashing · CUB200-2011, 48-bit | mAP (%): 26.57 -> 27.13 | Helped |
| Add cluster weights to part-oriented DenseCL (ablation) | Asymmetric augmented SSL hashing · CUB200-2011, 48-bit | mAP (%): 26.57 -> 28.32 | Helped |
| Add self-consistent hash code learning to asymmetric augmented SSL and unweighted part-oriented DenseCL (ablation) | Unsupervised hashing · CUB200-2011, 48-bit | mAP (%): 27.13 -> 28.85 | Helped |
| Add self-consistent hash code learning to asymmetric augmented SSL and weighted part-oriented DenseCL (ablation) | Unsupervised hashing · CUB200-2011, 48-bit | mAP (%): 28.32 -> 29.19 | Helped |
| Use only RandomCrop rather than traditional contrastive-learning augmentations (ablation) | Unsupervised hashing · CUB200-2011, 48-bit | mAP (%): 15.36 -> 23.57 | Helped |
| Remove GaussianBlur from traditional contrastive-learning augmentations (ablation) | Unsupervised hashing · CUB200-2011, 48-bit | mAP (%): 15.36 -> 16.69 | Helped |
Surprising findings
- Traditional contrastive learning with strong augmentations gave the lowest unsupervised retrieval accuracy on CUB200-2011, while removing augmentations steadily improved it, with random crop alone performing best.
- Applying ColorJitter or ElasticTransform to negative views improved retrieval, even though these augmentations hurt when applied to positive views.
Limitations
- Retrieval accuracy on fine-grained datasets remains low in absolute terms (e.g., 29.19% mAP on CUB at 48 bits).
- Failure cases occur when fine differences must be discerned under varying viewpoints.
- Number of clusters K must be tuned; larger K causes overparameterization and reduced gains.
Experimental setup
Datasets: CUB200-2011 Oxford Flowers Stanford Dogs Stanford Cars Food101
Benchmarks: CUB200-2011 Oxford Flowers Stanford Dogs Stanford Cars Food101
Metrics: mAP
Baselines: UnsupervisedGreedyHash MLS3RUDH BiHalf CIBHash SPQ MeCoQ SDC
Builds on: SimCLR DenseCL self-consistency principle Fisher Vector
Closest papers by meaning
- Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization (ICML 2025 · similarity 0.89)
- Deep Unsupervised Hashing via External Guidance (ICML 2025 · similarity 0.87)
- Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.86)
- Characteristics Matching Based Hash Codes Generation for Efficient Fine-grained Image Retrieval (CVPR 2024 · similarity 0.86)
- AutoSSVH: Exploring Automated Frame Sampling for Efficient Self-Supervised Video Hashing (CVPR 2025 · similarity 0.85)
- Efficient Self-Supervised Video Hashing with Selective State Spaces (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.