AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

TransFER: Learning Relation-Aware Facial Expression Representations With Transformers

Fanglei Xue, Qiangchang Wang, Guodong Guo

ICCV 2021 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 352 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 local CNNs (multiple LANet branches) to ViT baseline (ablation)TransFER · RAF-DB / AffectNetaccuracy: 89.93 -> 90.03 (RAF-DB), 65.63 -> 65.74 (AffectNet)No clear effect
add MAD to local CNNs (ablation)TransFER · RAF-DBaccuracy: 90.03 -> 90.35Improved
add MAD to local CNNs (ablation)TransFER · AffectNetmean class accuracy: 65.74 -> 65.94Improved
add MSAD (MAD in Transformer self-attention heads) (ablation)TransFER · RAF-DBaccuracy: 90.35 -> 90.91Improved
add MSAD (MAD in Transformer self-attention heads) (ablation)TransFER · AffectNetmean class accuracy: 65.94 -> 66.23Improved
use IR-50 stage 3 instead of stage 2 as stem output (ablation)TransFER · RAF-DBaccuracy: 84.94 -> 90.91Improved
use stage 3 instead of stage 4 of IR-50 (ablation)TransFER · RAF-DBaccuracy: 90.32 -> 90.91Improved
set number of local branches B to 2 rather than a small value (ablation)TransFER · RAF-DBaccuracy: 89.60 -> 90.91Improved
raise MAD drop rate p1 from 0.4 to 0.6 (p2=0.3) (ablation)TransFER · RAF-DBaccuracy: 89.28 -> 90.91Improved
raise MSAD drop rate p2 from 0.1 to 0.3 (p1=0.6), then to 0.5 (ablation)TransFER · RAF-DBaccuracy: 89.80 -> 90.91 -> 89.24Mixed
replace Dropout with MAD (ablation)TransFER · RAF-DBaccuracy: 90.35 -> 90.91Improved
replace Drop Block with MAD (ablation)TransFER · RAF-DBaccuracy: 90.25 -> 90.91Improved

Later papers that compared with this method

1 later accepted paper reports transfer as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.

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