Research map / Label-efficient and robust learning / Recognition and long-tail
Hard-Aware Deeply Cascaded Embedding
Yuhui Yuan, Kuiyuan Yang, Chao Zhang
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 hard example mining (single model) (ablation) | Contrastive (128-d) · CARS196 (ablation, no bbox) | Recall@1: 56.0 -> 67.6 | Improved |
| train on the target dataset instead of pretrained features only (ablation) | GoogLeNet · CARS196 (ablation, no bbox) | Recall@1: 40.5 -> 56.0 | Improved |
| replace single-model hard mining with cascaded hard-aware mining (same architecture) (ablation) | Hard + Contrastive · CARS196 (ablation, no bbox) | Recall@1: 67.6 -> 71.4 | Improved |
| concatenate features of all three cascaded models (ablation) | HDC + Contrastive-3 · CARS196 (ablation, no bbox) | Recall@1: 71.4 -> 73.7 | Improved |
| use deeper cascaded sub-model (model 1 to 2) (ablation) | HDC · CARS196 (ablation, no bbox) | Recall@1: 41.9 -> 58.0 | Improved |
| use deeper cascaded sub-model (model 2 to 3) (ablation) | HDC · CARS196 (ablation, no bbox) | Recall@1: 58.0 -> 71.4 | Improved |
| add cascade over single-model hard mining (ablation) | Hard + Contrastive · CARS196 (ablation, no bbox) | LDA score: 1.99 -> 2.50 | Improved |
| concatenate features of all three cascaded models (ablation) | HDC + Contrastive-3 · CUB-200-2011 (ablation, with bbox) | Recall@1: 58.5 -> 60.7 | Improved |
| use deeper cascaded sub-model (model 1 to 3) (ablation) | HDC · CUB-200-2011 (ablation, with bbox) | Recall@1: 43.4 -> 58.5 | Improved |
| HDC vs PDDM + Quadruplet with bounding boxes | HDC + Contrastive (384-d, bbox) · CARS196 (with bbox) | Recall@1: 57.4 -> 83.8 | Improved |
| HDC vs Npairs with multiple crops | HDC + Contrastive (384-d, multi-crop) · Stanford Online Products | Recall@1: 67.7 -> 70.1 | Improved |
| HDC vs previous best on Stanford Online Products | HDC + Contrastive (384-d) · Stanford Online Products | Recall@1: 65.5 -> 69.5 | Improved |
Closest papers by meaning
- Hardness-Aware Deep Metric Learning (CVPR 2019 · similarity 0.85)
- LoOp: Looking for Optimal Hard Negative Embeddings for Deep Metric Learning (ICCV 2021 · similarity 0.85)
- Smart Mining for Deep Metric Learning (ICCV 2017 · similarity 0.85)
- An Adversarial Approach to Hard Triplet Generation (ECCV 2018 · similarity 0.84)
- Hard negative examples are hard, but useful (ECCV 2020 · similarity 0.84)
- Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem (ICLR 2024 · 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.
