Research map / Label-efficient and robust learning / Recognition and long-tail
Significance of Softmax-based Features in Comparison to Distance Metric Learning-based Features
Shota Horiguchi, Daiki Ikami, Kiyoharu Aizawa
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 softmax-trained features with PCA and L2 normalization instead of DML (clustering loss) | GoogLeNet · CUB test (64-d) | Recall@1: 48.2 -> 51.1 | Improved |
| use softmax-trained features with PCA and L2 instead of clustering loss | GoogLeNet · CAR test (64-d) | Recall@1: 58.1 -> 69.4 | Improved |
| use softmax-trained features instead of DML on Online Products | GoogLeNet · OP test (64-d) | Recall@1: 67.0 -> 62.4 | Hurt |
| add PCA instead of random projection for dimension reduction (ablation) | softmax features + L2 · CUB test (64-d) | Recall@1: 47.5 -> 51.1 | Improved |
| add PCA instead of FC reduction layer (ablation) | softmax features + L2 · CUB test (64-d) | NMI: 59.1 -> 60.8 | Improved |
| add L2 normalization to FCR1 features (ablation) | FCR 1 · CUB NMI (1024-d) | NMI: 61.1 -> 62.0 | Improved |
| add L2 normalization to FCR2 features (ablation) | FCR 2 · CAR NMI (1024-d) | NMI: 58.7 -> 60.0 | Improved |
| use probability invariant shift distance instead of plain Euclidean (ablation) | FCR 1 · CUB NMI (1024-d) | NMI: 61.1 -> 59.3 | Hurt |
| use probability invariant shift distance (ablation) | FCR 2 · CAR NMI (512-d) | NMI: 61.4 -> 60.7 | Hurt |
| use L2 normalization instead of probability invariant shift (ablation) | FCR 2 · CAR NMI (256-d) | NMI: 61.3 -> 62.5 | Improved |
| increase training set size (samples per class) (ablation) | softmax features vs DML · CUB and CAR (5% to 100% data) | NMI and Recall@K | Improved |
| increase feature dimensionality to 1024 (ablation) | FCR2 and lifted struct · CAR NMI | NMI | Hurt |
Closest papers by meaning
- Large-Scale Distance Metric Learning With Uncertainty (CVPR 2018 · similarity 0.86)
- Distance Metric Learning with Joint Representation Diversification (ICML 2020 · similarity 0.86)
- Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.85)
- DiVA: Diverse Visual Feature Aggregation for Deep Metric Learning (ECCV 2020 · similarity 0.85)
- Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation (AAAI 2024 · similarity 0.83)
- Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.82)
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.
