Diverse hand gesture recognition dataset

• Published in: Multimedia tools and applications Vol. 83; no. 17; pp. 50245 - 50267
• Main Authors: Mohammadi, Zahra, Akhavanpour, Alireza, Rastgoo, Razieh, Sabokrou, Mohammad
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2024; Springer Nature B.V
• Subjects: Annotations; Computer Communication Networks; Computer Science; Computer vision; Data Structures and Information Theory; Datasets; Deep learning; Gesture recognition; Image reconstruction; Multimedia Information Systems; Special Purpose and Application-Based Systems; Deep learning; Computer vision; Hand Pose Estimation (HPE); Dataset; Hand Gesture Recognition (HGR)
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17268-8

Hand Gesture Recognition(HGR) is a challenging computer vision task. Recently, by taking advantages of deep learning-based models, HGR methods have achieved outstanding results and outperformed state-of-the-art alternatives by a high margin. However, the performance of deep learning-based models is highly dependent on the data. A large amount of data is required to train deep learning-based models. While there are some widely-used datasets in HGR, these datasets lack diverse gestures in real-world situations. To this end, we propose a hand gesture dataset (Dataset will be publicly available after paper publication.), including diverse gestures with more sample numbers per gesture class. Furthermore, we provide hand annotations, including a hand bounding box, 3D hand keypoints, and gesture label per sample. The proposed dataset aims to provide a benchmark for research works to tackle real-world situations. The dataset samples are recorded in a real-world background with high complexity and diversity. To be more realistic, the proposed dataset does not include any pre-processing step. All of the samples in this dataset are pure and real. This configuration makes room to underpin future research works in a real-world situation and develop gesture recognition models in an unrestricted environment. Overall, our dataset outperforms in terms of diversity, number of subjects, number of samples per gesture class, and use of real data. Finally, different analysis on the existing state-of-the-art models in HGR, HPE, hand recovery, and hand reconstruction were performed and reported. Our implementation is available at https://github.com/smohammadi96/Diverse_hand_gesture_dataset/blob/main/README.md .