AI Paper Atlas

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

Operational Open-Set Recognition and PostMax Refinement

Steve Cruz, Ryan Rabinowitz, Manuel Günther, Terrance E. Boult

ECCV 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Loss design and classifier theory · 13 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
normalize maximum logit by deep feature magnitude and map with a generalized Pareto distribution (PostMax) instead of raw MSP or MaxLogitHiera-H on ILSVRC2012 · OpenImage-OOOSA: 0.814 -> 0.864 vs MSPImproved
PostMax versus MSP on a hard fine-grained unknown set (ablation)Hiera-H and ConvNeXtV2-H · 21K-P Hard test, 21K-P Easy surrogateOOSA: 0.634 vs MSP 0.673 (Hiera-H)Hurt
swap validation knowns to ILSVRC2012 val and unknown surrogate to 21K-P Easy (ablation)PostMax · iNaturalist, OpenImage-OOOSA: 0.825 -> 0.863 (iNaturalist, Hiera-H)No clear effect
operate on ConvNeXtV2-H instead of Hiera-H (ablation)PostMax · iNaturalistOOSA: 0.825 -> 0.812No clear effect
divide logits by deep feature magnitude, since unknowns have larger feature normsPostMax · ViT-H feature analysis on class 990mean L2 norm of features: known 3.48 vs unknown 5.50 mean L2 normImproved
use PostMax with GPD probabilities versus NNGuide and SCALE OOD baselinesHiera-H · 21K-P EasyOOSA: 0.818 vs NNGuide 0.744 and SCALE 0.568Improved

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