AI Paper Atlas

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

Sparsity-Inducing Divergence Losses for Biometric Verification

Dimitrios Koutsianos, Ladislav Mosner, Yannis Panagakis, Themos Stafylakis

ECCV 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Fair and quality-aware face recognition · 0 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
use probabilistic reference-measure margin (Q-Margin) vs ArcFacealpha-divergence loss · IJB-B TAR@1e-5 (WF42M, R-100)TAR: 92.26 -> 93.67Improved
use Q-Margin vs ArcFacealpha-divergence loss · IJB-C TAR@1e-5TAR: 95.70 -> 96.35Improved
use Q-Margin vs ArcFacealpha-divergence loss · IJB-C TAR@1e-4TAR: 97.54 -> 97.78Improved
use Q-Margin vs PartialFC (reported)Q-Margin · IJB-C TAR@1e-4TAR: 97.82 -> 97.78No clear effect
increase margin from 0.2 to 0.5 (ablation)Q-Margin a1.25 s32 · IJB-C TAR@1e-5TAR: 96.35 -> 96.00Hurt
increase alpha from 1.25 to 1.5 (ablation)Q-Margin s32 m0.2 · IJB-B TAR@1e-5TAR: 93.38 -> 92.15Hurt
raise scale from 10 to 35 at alpha 1.25 (ablation)Q-Margin m0.2 · IJB-B TAR@1e-5TAR: 92.95 -> 93.67Improved
use SparseMax (alpha=2) instead of Q-Marginalpha-divergence loss · IJB-C TAR@1e-5TAR: 96.35 -> 95.28Hurt
restrict bisection to top 5% logits (ablation)Q-Margin · WebFace42M throughputsamples/s: 1466.45 -> 1900.26Improved
use higher alpha (1.5/1.75) for speaker verificationQ-Margin · VoxCeleb1-H TAR@1e-4TAR: 72.08 -> 74.60 vs ArcFaceImproved
use Q-Margin alpha=1.75 vs alpha=1.5 (ablation)Q-Margin · VoxCeleb1-HTAR: 87.02 vs 87.91 at 1e-3; 74.60 vs 72.96 at 1e-4Mixed
add angular shift on top of probabilistic margin (ablation)Q-Margin · MS1MV3 and WebFace42M (preliminary)TARNo clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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