Pose-Based Temporal Convolutional Networks for Isolated Indian Sign Language Word Recognition

Tatigunta Bhavi Teja Reddy, Vidhya Kamakshi


Abstract
This paper presents a lightweight and efficient baseline for isolated Indian Sign Language (ISL) word recognition developed forthe WSLP-AACL-2025 Shared Task. Wepropose a two-stage framework combiningskeletal landmark extraction via MediaPipeHolistic with a Temporal Convolutional Network (TCN) for temporal sequence classification. The system processes pose-basedinput sequences instead of raw video, significantly reducing computation and memorycosts. Trained on the WSLP-AACL-2025dataset containing 4,398 isolated sign videosacross 4,361 word classes, our model achieves54% top-1 and 78% top-5 accuracy.
Anthology ID:
2025.wslp-main.8
Volume:
Proceedings of the Workshop on Sign Language Processing (WSLP)
Month:
December
Year:
2025
Address:
IIT Bombay, Mumbai, India (Co-located with IJCNLP–AACL 2025)
Editors:
Mohammed Hasanuzzaman, Facundo Manuel Quiroga, Ashutosh Modi, Sabyasachi Kamila, Keren Artiaga, Abhinav Joshi, Sanjeet Singh
Venues:
WSLP | WS
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
47–50
Language:
URL:
https://preview.aclanthology.org/ingest-ijcnlp-aacl/2025.wslp-main.8/
DOI:
Bibkey:
Cite (ACL):
Tatigunta Bhavi Teja Reddy and Vidhya Kamakshi. 2025. Pose-Based Temporal Convolutional Networks for Isolated Indian Sign Language Word Recognition. In Proceedings of the Workshop on Sign Language Processing (WSLP), pages 47–50, IIT Bombay, Mumbai, India (Co-located with IJCNLP–AACL 2025). Association for Computational Linguistics.
Cite (Informal):
Pose-Based Temporal Convolutional Networks for Isolated Indian Sign Language Word Recognition (Reddy & Kamakshi, WSLP 2025)
Copy Citation:
PDF:
https://preview.aclanthology.org/ingest-ijcnlp-aacl/2025.wslp-main.8.pdf