Research map / Label-efficient and robust learning / Recognition and long-tail
Hashing with Binary Matrix Pursuit
Fatih Cakir, Kun He, Stan Sclaroff
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add regressed bit weights (weighted Hamming) over constant step sizes (ablation) | HBMP · NUSWIDE nus-1 (16 bits) | mAP@5K: 0.675 -> 0.757 | Improved |
| add regressed bit weights over constant step sizes (ablation) | HBMP · NUSWIDE nus-2 (16 bits) | mAP@50K: 0.746 -> 0.804 | Improved |
| add regressed bit weights over constant step sizes (ablation) | HBMP · NUSWIDE multi-level (16 bits) | NDCG: 0.905 -> 0.914 | Improved |
| add regressed bit weights over constant step sizes, long codes (ablation) | HBMP · 22K LabelMe multi-level (96 bits) | NDCG: 0.865 -> 0.917 | Improved |
| add regressed bit weights over constant step sizes, short codes (ablation) | HBMP · 22K LabelMe multi-level (16 bits) | NDCG: 0.837 -> 0.823 | Hurt |
| add regressed bit weights over constant step sizes (ablation) | HBMP · CIFAR-10 cifar-2 (16 bits) | mAP: 0.941 -> 0.942 | No clear effect |
| add regressed bit weights over constant step sizes (ablation) | HBMP · ImageNet100 (64 bits) | mAP@1K: 0.751 -> 0.742 | No clear effect |
| infer codes on classes instead of training instances (ablation) | two-stage hashing · 10-class synthetic affinity | mAP: difference can be as large as 0.40 | Improved |
| infer codes on classes instead of instances (residual convergence) (ablation) | two-stage hashing · 10-class synthetic affinity | residual norm: 40 bit codes reconstruct class affinity matrix | Improved |
| increase hash function capacity from single layer on GIST to fine-tuned VGG-F (ablation) | hash function learning · CIFAR-10 training set, 32-bit codes | unmatched bit ratio: above 0.25 -> well below 10% | Improved |
| two-stage HBMP versus MIHash | HBMP · CIFAR-10 cifar-1 (12 bits) | mAP: 0.738 -> 0.799 | Improved |
| two-stage HBMP versus MIHash | HBMP · ImageNet100 (64 bits) | mAP@1K: 0.694 -> 0.742 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MIHash — MIHash: Online Hashing With Mutual Information (ICCV 2017) · better on 13 of 18 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.91)
- Greedy Hash: Towards Fast Optimization for Accurate Hash Coding in CNN (NeurIPS 2018 · similarity 0.90)
- Optimizing affinity-based binary hashing using auxiliary coordinates (NeurIPS 2016 · similarity 0.90)
- Maximum-Margin Hamming Hashing (ICCV 2019 · similarity 0.90)
- Discrete Image Hashing Using Large Weakly Annotated Photo Collections (AAAI 2016 · similarity 0.90)
- MIHash: Online Hashing With Mutual Information (ICCV 2017 · similarity 0.89)
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.
