---
license: cc-by-nc-4.0
language:
  - en
tags:
  - sign-language
  - fingerspelling
  - indian-sign-language
  - video-classification
  - ctc
pipeline_tag: video-classification
---

# ISL Fingerspelling — model weights

Checkpoints for the ICPR 2026 paper **Dense Frame Annotations for Low-Resource ISL
Fingerspelling Recognition**
([Springer](https://dl.acm.org/doi/10.1007/978-3-032-31930-2_18)).

Inference code: <https://github.com/Kirandevraj/ISL-Fingerspelling-Dense>
Dataset: <https://huggingface.co/datasets/kirandevraj/ISL-Fingerspelling>

These files are downloaded automatically on first run — you do not need to fetch them
by hand:

```bash
git clone https://github.com/Kirandevraj/ISL-Fingerspelling-Dense
cd ISL-Fingerspelling-Dense
pip install -r requirements.txt
bash demo.sh
```

## Files

| File | Size | What it is |
|---|---|---|
| `recognition_standard.pt` | 57.5 MB | ResNet-18 + 2-layer BiLSTM + CTC head, trained with CTC **and** frame-level cross-entropy. Standard split. |
| `recognition_signer.pt` | 57.5 MB | Same architecture, signer-independent split. |
| `frame_classifier_standard.pt` | 44.8 MB | Stage-1 ResNet-18 frame classifier, 27 classes. Drives localization. |
| `frame_classifier_signer.pt` | 44.8 MB | Same, signer-independent split. |

Recognition is 14.3M parameters. Input is 224×224 RGB frames from **signer-cropped**
video; the CTC vocabulary is 28 (blank, space, a–z), while the standalone frame
classifier has 27 outputs (space, a–z — no blank).

## Citation

```bibtex
@inproceedings{islfs2026,
  title     = {Dense Frame Annotations for Low-Resource ISL Fingerspelling Recognition},
  booktitle = {International Conference on Pattern Recognition (ICPR)},
  year      = {2026},
  doi       = {10.1007/978-3-032-31930-2_18}
}
```