# Learning from What is Already Out There: Few-shot Sign Language Recognition with Online Dictionaries

**Authors:** Matyáš Boháček, Marek Hrúz  
**arXiv:** [2301.03769](https://arxiv.org/abs/2301.03769)  
**Date:** 2023-01-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.03769v1/

## Key points

- Authors Matyáš Boháček and Marek Hrúz address challenges in current sign language recognition models:
- Models require large training datasets of laboratory-like videos, which are difficult and costly to collect.
- Limited availability of publicly accessible systems, especially for less-populated sign languages.
- Proposal to overcome limitations and democratize technology:
- Utilize online text-to-video dictionaries containing annotated data on various attributes and sign languages.
- Introduce UWB-SL-Wild few-shot dataset sourced from dictionary-scraped videos to reflect actual distribution of online sign language data.
- Select glosses overlapping with existing datasets like WLASL100 and ASLLVD for transfer learning experiments.
- Novel approach to training sign language recognition models in a few-shot scenario.
- State-of-the-art results on ASLLVD-Skeleton and ASLLVD-Skeleton-20 datasets with top-1 accuracy rates of $30.97%$ and $95.45%$, respectively.
- Contribution to advancing sign language recognition technology:
- Addressing challenges related to training data availability.
- Making technology more inclusive across diverse linguistic communities.

## Summary

In their paper titled "Learning from What is Already Out There: Few-shot Sign Language Recognition with Online Dictionaries," authors Matyáš Boháček and Marek Hrúz address the challenges faced by current sign language recognition models. These models typically require large training datasets of laboratory-like videos, which can be difficult and costly to collect. This leads to a limited availability of publicly accessible systems, especially for less-populated sign languages. To overcome these limitations and democratize the technology, the authors propose utilizing online text-to-video dictionaries that contain annotated data on various attributes and sign languages. In this study, the researchers introduce the UWB-SL-Wild few-shot dataset sourced from dictionary-scraped videos. This dataset reflects the actual distribution and characteristics of online sign language data, providing a valuable resource for training recognition models in a few-shot fashion. By selecting glosses that overlap with existing datasets like WLASL100 and ASLLVD, the authors enable transfer learning experiments and facilitate comparisons between different datasets. Additionally, the paper presents a novel approach to training sign language recognition models in a few-shot scenario. The proposed method yields state-of-the-art results on ASLLVD-Skeleton and ASLLVD-Skeleton-20 datasets with impressive top-1 accuracy rates of $30.97~\%$ and $95.45~\%$, respectively. These results demonstrate the effectiveness of leveraging online dictionaries for training sign language recognition models and highlight the potential for broader accessibility and localization capabilities in this field. Overall, this work significantly contributes to advancing sign language recognition technology by addressing challenges related to training data availability and making it more inclusive across diverse linguistic communities. has been greatly improved through , utilizing  as a valuable . This approach also has the potential to enhance  capabilities and make sign language recognition more accessible for all.