English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Isolated Sign Language Recognition (SLR) has mostly been applied on datasetscontaining signs executed slowly and clearly by a limited group of signers. Inreal-world scenarios, however, we are met with challenging visual conditions,coarticulated signing, small datasets, and the need for signer independentmodels. To tackle this difficult problem, we require a robust feature extractorto process the sign language videos. One could expect human pose estimators tobe ideal candidates. However, due to a domain mismatch with their training setsand challenging poses in sign language, they lack robustness on sign languagedata and image-based models often still outperform keypoint-based models.Furthermore, whereas the common practice of transfer learning with image-basedmodels yields even higher accuracy, keypoint-based models are typically trainedfrom scratch on every SLR dataset. These factors limit their usefulness forSLR. From the existing literature, it is also not clear which, if any, poseestimator performs best for SLR. We compare the three most popular poseestimators for SLR: OpenPose, MMPose and MediaPipe. We show that throughkeypoint normalization, missing keypoint imputation, and learning a poseembedding, we can obtain significantly better results and enable transferlearning. We show that keypoint-based embeddings contain cross-lingualfeatures: they can transfer between sign languages and achieve competitiveperformance even when fine-tuning only the classifier layer of an SLR model ona target sign language. We furthermore achieve better performance usingfine-tuned transferred embeddings than models trained only on the target signlanguage. The embeddings can also be learned in a multilingual fashion. Theapplication of these embeddings could prove particularly useful for lowresource sign languages in the future.

Towards the extraction of robust sign embeddings for low resource sign language recognition