AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Sign Language Transformers: Joint End-to-End Sign Language Recognition and Translation

Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, Richard Bowden

CVPR 2020 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 105 citations (OpenAlex)

This paper tried a common change: Add batch / layer normalization

add batch normalization and ReLU to pretrained sign-recognition CNN features · PHOENIX-2014T test · WER: 33.39 -> 27.62 Improved

What 431 accepted papers report for this change

See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.

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
jointly train sign recognition and translation with CTC gloss supervisionSign Language Transformers · PHOENIX-2014T testBLEU-4: 9.58 -> 20.17Improved
replace the RNN-based text-to-text system with a transformerGloss2Text · PHOENIX-2014T testBLEU-4: 19.26 -> 24.54Improved
add batch normalization and ReLU to pretrained sign-recognition CNN features (ablation)Sign2Gloss · PHOENIX-2014T testWER: 33.39 -> 27.62Improved
use pretrained sign-recognition CNN features instead of EfficientNet-B7 (ablation)Sign2Gloss · PHOENIX-2014T testWER: 47.96 -> 33.39Improved
increase the SLRT depth from 1 to 3 layers (ablation)Sign2Gloss · PHOENIX-2014T testWER: 29.90 -> 24.59Improved
increase the SLRT depth from 3 to 6 layers (ablation)Sign2Gloss · PHOENIX-2014T testWER: 24.59 -> 27.78Hurt
jointly train with equal recognition and translation loss weights instead of recognition-only training (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testWER: 24.59 -> 33.75Hurt
jointly train with equal recognition and translation loss weights instead of translation-only training (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testBLEU-4: 20.17 -> 21.22Improved
increase recognition loss weight from 1.0 to 5.0 while keeping translation weight at 1.0 (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testWER: 33.75 -> 24.49Improved
increase recognition loss weight from 1.0 to 5.0 while keeping translation weight at 1.0 (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testBLEU-4: 21.22 -> 21.80Improved
increase recognition loss weight from 5.0 to 10.0 while keeping translation weight at 1.0 (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testBLEU-4: 21.80 -> 21.32Improved
increase recognition loss weight from 5.0 to 20.0 while keeping translation weight at 1.0 (ablation)Sign2(Gloss+Text) · PHOENIX-2014T testWER: 24.49 -> 25.73Hurt

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