Research map / Label-efficient and robust learning / Recognition and long-tail
Maximum-Margin Hamming Hashing
Rong Kang, Yue Cao, Mingsheng Long, Jianmin Wang, Philip S. Yu
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 quantization loss (ablation) | MMHH · NUS-WIDE, MS-COCO, CIFAR-10 | MAP@H<=2: 4.0%, 3.1% and 4.2% | Improved |
| use cosine distance instead of Euclidean distance as Hamming proxy (ablation) | MMHH · NUS-WIDE, MS-COCO, CIFAR-10 | MAP@H<=2: 16.2%, 12.2% and 22.8% | Improved |
| add semi-batch optimization (ablation) | MMHH, DCH, HashNet · ImageNet subsets (100-1000 classes) | MAP@H<=2 | Improved |
| more classes increase the gain from semi-batch optimization | MMHH · ImageNet-1000 vs ImageNet-100 | MAP@H<=2 | Improved |
| replace Cauchy loss (DCH) with max-margin t-distribution loss | DCH · NUS-WIDE 50% label noise | MAP drop: 0.06 -> 0.02 | Improved |
| replace sigmoid loss (HashNet) with max-margin t-distribution loss | HashNet · NUS-WIDE 50% label noise | MAP drop: 0.17 -> 0.02 | Improved |
| use max-margin t-distribution loss instead of HashNet sigmoid loss | HashNet · MS-COCO 48 bits | queries returning no images: 44% -> 13% | Improved |
| use max-margin loss instead of Cauchy loss to reduce loss on in-ball similar pairs | DCH · MS-COCO 48 bits | share of total loss from retrievable similar pairs: ~40% | Improved |
| hash lookup instead of linear scan | MMHH · MS-COCO | search time per query: 10.2 ms/query -> 0.9 ms/query | Improved |
| use deep features with shallow SDH vs end-to-end MMHH | MMHH vs SDH · NUS-WIDE, MS-COCO, CIFAR-10 | average MAP@H<=2: 17.5%, 19.0% and 23.5% | Improved |
| evaluate on unseen classes instead of seen classes | MMHH · MS-COCO unseen-class protocol | MAP | Hurt |
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.
- HashNet — HashNet: Deep Learning to Hash by Continuation (ICCV 2017) · better on 6 of 6 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 Cauchy Hashing for Hamming Space Retrieval (CVPR 2018 · similarity 0.93)
- Improved Hamming Distance Search Using Variable Length Substrings (CVPR 2016 · similarity 0.90)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.90)
- MIHash: Online Hashing With Mutual Information (ICCV 2017 · similarity 0.90)
- Structured Learning of Binary Codes with Column Generation for Optimizing Ranking Measures (IJCV 2017 · similarity 0.90)
- Hashing with Binary Matrix Pursuit (ECCV 2018 · similarity 0.90)
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.
