Continuous Fingerspelling Dataset for Indian Sign Language

WSLP @ AACL-IJCNLP 2025 ICPR 2026
R Kirandevraj1, Vinod K Kurmi2, Vinay P Namboodiri3, CV Jawahar1
1IIIT Hyderabad, India    2IISER Bhopal, India    3University of Bath, UK

We release a continuous fingerspelling resource for Indian Sign Language, comprising 1,308 video segments with aligned text. Building on this, we add character-level frame annotations for every letter and an automatically discovered set of 11,700+ additional segments, supporting research in fingerspelling recognition, detection, temporal localization, and sign generation.

Fingerspelling: "sulochana das"
Fingerspelling: "vinesh phogat"

Releases

ISL-FS — Continuous Fingerspelling Dataset WSLP @ AACL-IJCNLP 2025

The first continuous fingerspelling dataset for ISL: 1,308 segments from 499 ISH News videos with aligned word-level text. A ByT5-small baseline reaches 82.91% CER after fine-tuning.

1,308 Video Segments
70 min Total Duration
14,814 Characters
499 Source Videos
3 Signers
90.67% Validation Acc.

ISL-FS-Dense — Character-Level Frame Annotations & Auto-Discovered Segments ICPR 2026 · New

Extends ISL-FS with character-level frame boundaries for every letter in all 1,308 segments (99.1% expert-validated), and 11,700+ auto-discovered segments mined from the iSign corpus. Dense supervision reaches 4.87% CER, beating word-level baselines (16.4%) and BSL transfer (8.1%); the same model also yields 84.6% F1 fingerspelling detection without dedicated detection training.

125,661 Annotated Frames
14,682 Character Instances
38 Unique Characters
99.1% Expert Validation
11,700+ Auto-Discovered Segments
4.87% Recognition CER
84.6% Detection F1
~35 hr Annotation Effort

Annotation & Validation

Boundaries follow a transition-onset rule: marked at the frame where the hand begins moving toward the next letter shape. A professional ISL interpreter blindly re-transcribed 660 cropped letter segments; only 6 were genuine errors, giving 99.1% validated accuracy.

Recognition Results

Method Input Supervision Standard CER Signer-Indep CER
ByT5 RGB Word 85.9% 81.4%
Transpeller (BSL pretrained) RGB Word 8.1% 15.7%
ResNet-BiLSTM RGB Word 16.4% 37.4%
ResNet-BiLSTM RGB Frame + Word 4.87% 16.8%
MLP-BiLSTM Keypoint Frame + Word 5.1% 14.3%

Automatic Dataset Scaling

Two pipelines mine the iSign corpus: a multi-scale sliding window (7,833 segments) and detection-based matching against NER-extracted proper nouns (7,479 segments, 51.7% non-overlapping). Adding them to training improves signer-independent CER by 10% relative on RGB and 6% on keypoints.

Dataset Structure

ISL-FS (word-level)

ISL-Fingerspelling/
├── videos/                        # 1,308 MP4 video files
├── fingerspelling_annotations.csv # Segment annotations
├── localization_timestamps.csv    # Temporal localization in source videos
└── README.md

ISL-FS-Dense (frame-level + auto-discovered)

The ISL-FS-Dense release will include character-level frame boundaries for every letter in the 1,308 manually annotated segments, plus the 11,700+ automatically discovered segments with their matched transcripts. Dataset link will be added here once published.

Citation

If you use either release in your research, please cite the relevant paper(s):

ISL-FS (WSLP @ AACL-IJCNLP 2025)

@inproceedings{r-etal-2025-continuous,
  title     = {Continuous Fingerspelling Dataset for {I}ndian {S}ign {L}anguage},
  author    = {R, Kirandevraj and Kurmi, Vinod K. and Namboodiri, Vinay P. and Jawahar, C.V.},
  booktitle = {Proceedings of the Workshop on Sign Language Processing (WSLP)},
  publisher = {Association for Computational Linguistics},
  year      = {2025},
  pages     = {33--38},
  url       = {https://aclanthology.org/2025.wslp-main.6/},
  doi       = {10.18653/v1/2025.wslp-main.6}
}

ISL-FS-Dense (ICPR 2026)

@inproceedings{kirandevraj2026denseisl,
  title={Dense Frame Annotations for Low-Resource ISL Fingerspelling Recognition},
  author={Kirandevraj, R and Kurmi, Vinod K and Namboodiri, Vinay P and Jawahar, CV},
  booktitle={Proceedings of the International Conference on Pattern Recognition (ICPR)},
  year={2026}
}

Acknowledgments

Data sourced from publicly available ISH News YouTube videos. 407 of 499 videos overlap with the iSign dataset. Both datasets are released under CC BY-NC 4.0 license for research purposes only.