Research map / Label-efficient and robust learning / Recognition and long-tail
The Group Loss for Deep Metric Learning
Ismail Elezi, Sebastiano Vascon, Alessandro Torcinovich, Marcello Pelillo, Laura Leal-Taixé
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 |
|---|---|---|---|
| clamp negative similarities with ReLU instead of shifting (ablation) | Group Loss · CUB-200-2011 | Recall@1: 51 -> 64.3 | Improved |
| clamp negative similarities with ReLU instead of shifting (ablation) | Group Loss · Cars 196 | Recall@1 | No clear effect |
| add temperature scaling to the softmax priors (ablation) | Group Loss · CUB-200-2011 | Recall@1: +2.5pp | Improved |
| add temperature scaling to the softmax priors (ablation) | Group Loss · Cars 196 | Recall@1: +2pp | Improved |
| replace cross-entropy with KL divergence between predicted distributions of same-class samples (ablation) | Group Loss · CUB-200-2011 | Recall@1: 3pp lower | Hurt |
| replace cross-entropy with KL divergence between predicted distributions of same-class samples (ablation) | Group Loss · Cars 196 | Recall@1: 1.5pp lower | Hurt |
| tune hyperparameters with random search (ablation) | Group Loss Densenet161 · Cars 196 | Recall@1: 90.7 (2pp higher than default) | Improved |
| tune hyperparameters with random search (ablation) | Group Loss Densenet161 · Cars 196 | NMI: 77.6 (3pp higher than default) | Improved |
| use larger Densenet161 backbone instead of Densenet121 (ablation) | Group Loss · Stanford Online Products | Recall@1: 78.2 -> 80.3 | Improved |
| use larger Densenet161 backbone instead of Densenet121 (ablation) | Group Loss · CUB-200-2011 | Recall@1: 65.5 -> 64.7 | Hurt |
| use multiple centres (10 per class) instead of one for SoftTriple (ablation) | SoftTriple · Cars 196 | Recall@1: 83.1 -> 84.5 | Improved |
| vary number of anchors (0 to 4) and samples per class (5 to 10) (ablation) | Group Loss · Cars 196 | Recall@1: 2.1 percentage points between best and worst | No clear effect |
Closest papers by meaning
- The Group Loss++: A Deeper Look Into Group Loss for Deep Metric Learning (TPAMI 2022 · similarity 0.88)
- Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning (CVPR 2019 · similarity 0.85)
- Towards Improved and Interpretable Deep Metric Learning via Attentive Grouping (TPAMI 2022 · similarity 0.84)
- Improved Deep Metric Learning with Multi-class N-pair Loss Objective (NeurIPS 2016 · similarity 0.84)
- Moving in the Right Direction: A Regularization for Deep Metric Learning (CVPR 2020 · similarity 0.83)
- Mean Field Theory in Deep Metric Learning (ICLR 2024 · similarity 0.81)
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.
