Research map / Label-efficient and robust learning / Recognition and long-tail
Hashing with Mutual Information
Fatih Çakir, Kun He, Sarah Adel Bargal, 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 |
|---|---|---|---|
| use mutual information objective instead of DTSH triplet loss | hashing with VGG-F · CIFAR-10 (500 per class training) | mAP: 0.765 -> 0.791 (32 bits), 0.774 -> 0.816 (48 bits) | Improved |
| use mutual information objective instead of DTSH | hashing with VGG-F · NUSWIDE (500 per label training) | mAP@5K: 0.773 -> 0.773 (12 bits), 0.820 (DTSH 0.813) at 24 bits | Mixed |
| use mutual information objective with more training data | hashing with VGG-F · CIFAR-10 and NUSWIDE (full training set) | mAP: 0.926 -> 0.943 (CIFAR 48 bits), 0.756 -> 0.798 (NUSWIDE 16 bits) | Improved |
| use mutual information objective instead of HashNet | hashing with AlexNet · ImageNet100 | mAP@1K: 0.506 -> 0.569 (16 bits) | Improved |
| use mutual information objective with one-layer embedding on GIST | shallow hashing · 22K LabelMe | mAP: 0.369 -> 0.384 (16 bits), 0.582 -> 0.598 (64 bits) | Improved |
| learn hashing with MI objective (online, 20K examples) from LSH init | MIHash single-layer on fc7 · CIFAR-10 | mAP: 0.22 -> 0.68 | Improved |
| increase steepness gamma from 1 (ablation) | MIHash · CIFAR-10 (500 per class training), 32 bits | mAP: 0.791 -> 0.768 (5), 0.749 (10), 0.776 (20) | No clear effect |
| vary minibatch size M from 256 (ablation) | MIHash · CIFAR-10 (500 per class training), 32 bits | mAP: 0.791 -> 0.771 (64), 0.765 (128), 0.783 (512) | No clear effect |
| MI objective value correlates with mAP | random LSH-style hash mappings · CIFAR-10 | mAP: Pearson Correlation Coefficient score of 0.98 | 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.
- HashNet — HashNet: Deep Learning to Hash by Continuation (ICCV 2017) · better on 4 of 4 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
- MIHash: Online Hashing With Mutual Information (ICCV 2017 · similarity 0.92)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.91)
- Deep Hashing Network for Efficient Similarity Retrieval (AAAI 2016 · similarity 0.90)
- An ensemble diversity approach to supervised binary hashing (NeurIPS 2016 · similarity 0.90)
- Hashing as Tie-Aware Learning to Rank (CVPR 2018 · similarity 0.90)
- Maximum-Margin Hamming Hashing (ICCV 2019 · 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.
