nithinm2719/SignFlow-ASL-Citizen-SOTA

SignFlow: Real-Time Multi-Modal Sign Language Recognition

SignFlow is an end-to-end, on-device sign language recognition system evaluated on the complete official ASL Citizen benchmark (2,731 glosses, 32,941 test clips across 11 unseen signers).

• Official Code & Reproduction: GitHub Repository
• Permanent Archival DOI: 10.5281/zenodo.24088419
• Zenodo Record: https://zenodo.org/records/24088419

Benchmark Results on Official ASL Citizen Test Split (2,731 Glosses)

Model / Architecture	Modality	Top-1 Accuracy	Top-5 Accuracy	Top-10 Accuracy	Notes
SignFlow (12-stream hybrid)	Landmarks + RGB	87.03%	98.27%	98.88%	#1 SOTA (Log-linear fusion fitted on Val)
SignFlow (conv3 solo expert)	Landmarks only	85.15%	97.35%	—	Single model (No ensemble)
SignFlow (Pure Skeleton 7-stream)	Landmarks only	85.74%	97.66%	98.74%	4 transformers + 3 conv experts
SignFlow (Conv-only 3-stream)	Landmarks only	85.57%	97.56%	98.71%	3 dilated temporal conv experts
SignBart (CVPR 2025)	Skeleton	75.22%*	—	—	Reported on validation split
PoseFormer (2025)	Skeleton	75.10%	—	—	Prior benchmark best
I3D Official Baseline (NeurIPS 2023)	RGB Video	63.10%	83.50%	91.00%	Desai et al.
ST-GCN Official Baseline (NeurIPS 2023)	Skeleton Graph	59.52%	—	—	Desai et al.

Author & Citation

• Author: Nithin G (Independent Researcher)
• Contact: thangamnithin1980@gmail.com

@article{signflow2026,
title   = {SignFlow: Real-Time Multi-Modal Sign Language Recognition via a 12-Stream Hybrid Architecture on the Full-Vocabulary ASL Citizen Benchmark},
author  = {Nithin G},
year    = {2026},
doi     = {10.5281/zenodo.24088419},
url     = {https://doi.org/10.5281/zenodo.24088419}
 }