Research map / Label-efficient and robust learning / Recognition and long-tail
Permutation-Invariant Feature Restructuring for Correlation-Aware Image Set-Based Recognition
Xiaofeng Liu, Zhenhua Guo, Site Li, Lingsheng Kong, Ping Jia, Jane You, B.V.K. Vijaya Kumar
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 DFA inter-set module (l2) to RSA (ablation) | RSA (GoogleNet) · IJB-A 1:1 verification (GoogleNet) | TAR at FAR=0.001: 0.926 -> 0.937 | Improved |
| add DFA inter-set module (l2) to RSA (ablation) | RSA (ResNet50) · IJB-A 1:1 verification (ResNet50) | TAR at FAR=0.001: 0.943 -> 0.955 | Improved |
| add RSA inner-set module to DFA(l2) (ablation) | DFA(l2) (GoogleNet) · IJB-A 1:1 verification (GoogleNet) | TAR at FAR=0.001: 0.914 -> 0.937 | Improved |
| use l2 collaborative instead of l1 sparse coding in DFA (ablation) | PIFR (GoogleNet) · IJB-A 1:1 verification (GoogleNet) | TAR at FAR=0.001: 0.933 -> 0.937 | Improved |
| add symmetric dictionary learning (x2) (ablation) | PIFR(l2) (GoogleNet) · IJB-A 1:1 verification (GoogleNet) | TAR at FAR=0.001: 0.937 -> 0.938 | No clear effect |
| add DFA to RSA (ablation) | RSA (GoogleNet) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8558 -> 0.8604 | Improved |
| add Gaussian spatial similarity Delta to RSA (ablation) | RSA (GoogleNet) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8420 -> 0.8558 | Improved |
| use feature difference residual instead of neighbor feature as residue (ablation) | RSA (GoogleNet) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8388 -> 0.8558 | Improved |
| replace original non-local with RSA (ablation) | non-local (GoogleNet) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8251 -> 0.8558 | Improved |
| use embedded Gaussian affinity in every RSA block (ablation) | RSA (GoogleNet) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8556 -> 0.8558 | No clear effect |
| switch backbone from GoogleNet to ResNet50 | PIFR(l2) · Celebrity-1000 open-set identification, 800 subjects | rank-1 accuracy: 0.8604 -> 0.8713 | Improved |
| replace DAC with RSA | DAC (GoogleNet) · IJB-A 1:N identification (GoogleNet) | Rank-1 accuracy: 0.973 -> 0.984 | Improved |
Closest papers by meaning
- Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks (ICML 2019 · similarity 0.86)
- FaceCoresetNet: Differentiable Coresets for Face Set Recognition (AAAI 2024 · similarity 0.85)
- Simultaneous Feature and Sample Reduction for Image-Set Classification (AAAI 2016 · similarity 0.85)
- Reconstruction-Based Disentanglement for Pose-Invariant Face Recognition (ICCV 2017 · similarity 0.85)
- Dependency-aware Attention Control for Unconstrained Face Recognition with Image Sets (ECCV 2018 · similarity 0.85)
- Discriminative Covariance Oriented Representation Learning for Face Recognition With Image Sets (CVPR 2017 · 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.
