AI Paper Atlas

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

Self Supervision to Distillation for Long-Tailed Visual Recognition

Tianhao Li, Limin Wang, Gangshan Wu

ICCV 2021 · Label-efficient and robust learning › Recognition and long-tail · approach: Decoupled and calibrated long-tail training · 106 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
add instance-discrimination self-supervision to feature learning (ablation)CE baseline · ImageNet-LTtop-1 accuracy: 46.3 -> 48.9Improved
add self-supervision to teacher training (ablation)SSD · ImageNet-LTtop-1 accuracy: 54.4 -> 56.0Improved
add 1.5x training schedule (ablation)CE · ImageNet-LTtop-1 accuracy: 45.1 -> 46.3Improved
add 1.5x training schedule (ablation)LWS · ImageNet-LTtop-1 accuracy: 50.7 -> 52.1Improved
use soft-label self-distillation student instead of teacher (ablation)SSD · ImageNet-LTtop-1 accuracy: 54.1 -> 55.7Improved
use hybrid hard+soft supervision (two classifiers) instead of single soft-only (ablation)SSD · ImageNet-LTtop-1 accuracy: 55.1 -> 55.7Improved
use two-classifier hybrid supervision instead of coupled distillation (ablation)SSD · ImageNet-LTtop-1 accuracy: 53.2 -> 55.7Improved
fine-tune hard-label student with LWS (stage IV) (ablation)SSD · ImageNet-LTtop-1 accuracy: 0.3%Improved
add soft-label student to hard-label student (ablation)SSD · ImageNet-LTtop-1 accuracy: 51.6 -> 55.7Improved
replace LWS* with SSD (1.5x)LWS* · ImageNet-LTtop-1 accuracy: 52.1 -> 56.0Improved
replace LWS* with SSDLWS* · CIFAR100-LT (imbalance 100)top-1 accuracy: 42.3 -> 46.0Improved
replace LWS with SSDLWS‡* · iNaturalist 2018 (1x)top-1 accuracy: 66.6 -> 69.3Improved

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