Indian Sign Language Recognition with Landmark Features: Accuracy–Efficiency Tradeoffs Across Classical and Deep Models

Abstract

Indian Sign Language (ISL) recognition remains challenging for school-scale and resource-constrained deployments due to limited datasets and the computational cost of modern deep learning systems. This study investigates isolated ISL word recognition using a school-collected dataset of 67 signs and 1,072 videos recorded with four native signers. Hand and upper-body landmarks are extracted using MediaPipe and converted into spatial and temporal features for classification using XGBoost, a CNN–LSTM model, and a compact Transformer. On the held-out test set, the Transformer achieves 73.2% accuracy, while XGBoost achieves 71.0% accuracy with substantially lower computational cost. XGBoost requires approximately 15 minutes of training and achieves 45 ms inference latency, compared with 120 minutes of training and 180 ms inference latency for the Transformer. These results highlight an accuracy–efficiency tradeoff relevant to resource-constrained assistive technology deployments. The study also evaluates cross-validation stability and feature ablations, showing the importance of temporal features and data augmentation. The dataset contains 67 words across six categories and is publicly available for research use. The associated code is available through the project's GitHub repository. Important This description is based on the content of your actual paper, including its dataset, models, results, and availability claims. Dataset: https://www.kaggle.com/datasets/bugruster/isl-buggy-arn089Code: https://github.com/BugRuster/MJRIDK