Research map / Detection and segmentation / CNN features
Learning Surrogates via Deep Embedding
Yash Patel, Tomáš Hodaň, Jiří Matas
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 |
|---|---|---|---|
| post-tune with LS-ED instead of cross-entropy only | TPS-ResNet-BiLSTM-Attn · IIIT-5K | accuracy: 84.300 -> 86.300 | Improved |
| post-tune with LS-ED instead of cross-entropy only | ResNet-BiLSTM-Attn (no TPS) · ICDAR'15 | accuracy: 78.520 -> 78.410 | Hurt |
| post-tune with LS-ED | TPS-ResNet-BiLSTM-Attn · SVT | accuracy: 87.172 -> 86.708 | Hurt |
| post-tune with LS-ED | TPS-ResNet-BiLSTM-Attn · CUTE | accuracy: 73.171 -> 74.216 | Improved |
| post-tune with LS-ED | ResNet-BiLSTM-Attn (no TPS) · ICDAR'13 total edit distance | TED: 260 -> 157 | Improved |
| post-tune with LS-ED | ResNet-BiLSTM-Attn (no TPS) · ICDAR'15 | accuracy: 71.452 -> 74.655 | Improved |
| train the surrogate on random generator data only (global approximation) (ablation) | LS-IoU · ICDAR'15 text detection | F1 score: 77.37% -> 74.81% | Hurt |
| train the surrogate on model samples only (local approximation) (ablation) | LS-IoU · ICDAR'15 text detection | recall: 71.21% -> 70.92% | Hurt |
| train the surrogate on model samples only (local approximation) (ablation) | LS-IoU · ICDAR'15 text detection | precision: 84.71% -> 86.60% | Improved |
| use local-global approximation instead of global only (ablation) | LS-IoU · ICDAR'15 text detection | F1 score: 74.81% -> 80.66% | Improved |
| use local-global approximation instead of local only (ablation) | LS-IoU · ICDAR'15 text detection | F1 score: 77.98% -> 80.66% | Improved |
| replace smooth-L1 regression loss with LS-IoU (local-global) in post-tuning | RRPN-ResNet-50 · ICDAR'15 text detection | F1 score: 77.37% -> 80.66% | Improved |
Closest papers by meaning
- Relational Surrogate Loss Learning (ICLR 2022 · similarity 0.88)
- Sampling Matters in Deep Embedding Learning (ICCV 2017 · similarity 0.85)
- Combating Label Noise with a General Surrogate Model for Sample Selection (IJCV 2024 · similarity 0.83)
- Divide and Conquer the Embedding Space for Metric Learning (CVPR 2019 · similarity 0.83)
- On the Surrogate Gap between Contrastive and Supervised Losses (ICML 2022 · similarity 0.83)
- Surrogate Prompt Learning: Towards Efficient and Diverse Prompt Learning for Vision-Language Models (ICML 2025 · similarity 0.83)
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.
