AI Paper Atlas

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

ECCV 2020 · Label-efficient and robust learning › Recognition and long-tail · approach: Loss design and classifier theory · 55 citations (OpenAlex)

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 changedWhereMeasuredEffect
clamp negative similarities with ReLU instead of shifting (ablation)Group Loss · CUB-200-2011Recall@1: 51 -> 64.3Improved
clamp negative similarities with ReLU instead of shifting (ablation)Group Loss · Cars 196Recall@1No clear effect
add temperature scaling to the softmax priors (ablation)Group Loss · CUB-200-2011Recall@1: +2.5ppImproved
add temperature scaling to the softmax priors (ablation)Group Loss · Cars 196Recall@1: +2ppImproved
replace cross-entropy with KL divergence between predicted distributions of same-class samples (ablation)Group Loss · CUB-200-2011Recall@1: 3pp lowerHurt
replace cross-entropy with KL divergence between predicted distributions of same-class samples (ablation)Group Loss · Cars 196Recall@1: 1.5pp lowerHurt
tune hyperparameters with random search (ablation)Group Loss Densenet161 · Cars 196Recall@1: 90.7 (2pp higher than default)Improved
tune hyperparameters with random search (ablation)Group Loss Densenet161 · Cars 196NMI: 77.6 (3pp higher than default)Improved
use larger Densenet161 backbone instead of Densenet121 (ablation)Group Loss · Stanford Online ProductsRecall@1: 78.2 -> 80.3Improved
use larger Densenet161 backbone instead of Densenet121 (ablation)Group Loss · CUB-200-2011Recall@1: 65.5 -> 64.7Hurt
use multiple centres (10 per class) instead of one for SoftTriple (ablation)SoftTriple · Cars 196Recall@1: 83.1 -> 84.5Improved
vary number of anchors (0 to 4) and samples per class (5 to 10) (ablation)Group Loss · Cars 196Recall@1: 2.1 percentage points between best and worstNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy