Exploring Pose-based Sign Language Translation: Ablation Studies and Attention Insights

TL;DR

This paper investigates how pose-based preprocessing techniques affect sign language translation performance using a transformer model, revealing that normalization, interpolation, and augmentation improve robustness and that a register token enhances accuracy.

Contribution

It introduces a comprehensive analysis of pose preprocessing effects on SLT and proposes adding a register token to improve model performance.

Findings

Preprocessing techniques significantly boost translation accuracy.

Adding a register token improves overall model performance.

Deep attention analysis reveals key behaviors in the transformer model.

Abstract

Sign Language Translation (SLT) has evolved significantly, moving from isolated recognition approaches to complex, continuous gloss-free translation systems. This paper explores the impact of pose-based data preprocessing techniques - normalization, interpolation, and augmentation - on SLT performance. We employ a transformer-based architecture, adapting a modified T5 encoder-decoder model to process pose representations. Through extensive ablation studies on YouTubeASL and How2Sign datasets, we analyze how different preprocessing strategies affect translation accuracy. Our results demonstrate that appropriate normalization, interpolation, and augmentation techniques can significantly improve model robustness and generalization abilities. Additionally, we provide a deep analysis of the model's attentions and reveal interesting behavior suggesting that adding a dedicated register token can improve overall model performance. We publish our code on our GitHub repository, including the preprocessed YouTubeASL data.