AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

A Metric Learning Reality Check

Kevin Musgrave, Serge Belongie, Ser-Nam Lim

ECCV 2020 · Label-efficient and robust learning › Recognition and long-tail · approach: Loss design and classifier theory · 502 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
fine-tune with metric learning loss versus pretrained embeddingsBN-Inception · CUB200MAP@R (concatenated 512-dim): 14.21 -> 26.53Improved
use ArcFace instead of classic contrastive lossBN-Inception · CUB200MAP@R (concatenated 512-dim): 26.53 -> 26.45No clear effect
use MultiSimilarity instead of classic contrastive lossBN-Inception · Cars196MAP@R (concatenated 512-dim): 24.89 -> 28.07Improved
use MultiSimilarity instead of classic contrastive lossBN-Inception · CUB200MAP@R (concatenated 512-dim): 26.53 -> 24.70Hurt
use ProxyNCA instead of triplet lossBN-Inception · SOPMAP@R (concatenated 512-dim): 43.37 -> 47.22Improved
add online MS miner to MultiSimilarity loss (ablation)BN-Inception · CUB200MAP@R (concatenated 512-dim): 24.70 -> 26.52Improved
add online MS miner to MultiSimilarity loss (ablation)BN-Inception · Cars196MAP@R (concatenated 512-dim): 28.07 -> 27.01Hurt
learn separate margin beta per class instead of a shared beta (ablation)BN-Inception · CUB200MAP@R (concatenated 512-dim): 23.09 -> 23.71Improved
learn separate margin beta per class instead of a shared beta (ablation)BN-Inception · Cars196MAP@R (concatenated 512-dim): 24.21 -> 23.11Hurt
concatenate four cross-validation models' 128-dim embeddings to 512-dim instead of averaging separate accuracies (ablation)Contrastive loss · CUB200MAP@R: 21.18 -> 26.53Improved
evaluate with MAP@R instead of Recall@1validation accuracy during training · metric stabilitylag-one autocorrelation: 0.73 -> 0.81Improved
freeze BatchNorm parameters of pretrained model (ablation)metric learning training · CUB200Recall@1: 2 pointImproved

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