정적 손 제스처 인식을 위한 딥러닝 모델 비교 연구: DenseNet vs ViT with HaGRID
A Comparative Study of Deep Learning Models for Static Hand Gesture Recognition: DenseNet vs. ViT with HaGRID
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
본 연구는 증강현실·가상현실 등 HCI 응용 분야에서 중요한 손 제스처 인식(HGR)을 위해 대규모 HaGRID v2 512px 데이터 셋을 활용하여 DenseNet-121과 Vision Transformer (ViT-B/16) 모델을 비교 평가하였다. DenseNet-121은 95.32%, ViT-B/16은 95.45%의 테스트 정확도를 기록하였으며, 특히 ViT-B/16은 시각적으로 유사한 제스처 간 오분류를 줄이는 데 효과적이었 다. 연구 결과는 트랜스포머 기반 모델이 그래픽 중심의 HCI 분야에서 실용적임을 보여준다.
Abstract
Hand Gesture Recognition (HGR) is essential for natural and intuitive interactions in fields like augmented reality, virtual reality, and mixed reality, significantly enhancing the user experience in human-computer interaction (HCI) applications. In this study, we present an extensive evaluation of static Hand Gesture Recognition models using the large-scale HaGRID v2 512px dataset, comprising 1,086,167 RGB images, covering 33 gesture classes along with a dedicated no_gesture category, from over 65,977 unique individuals. We systematically benchmark two state-of-the-art deep learning models: a lightweight convolutional neural network (DenseNet-121), trained from scratch, and a Vision Transformer (ViT-B/16) fine-tuned from ImageNet-21k pre-training. Under identical training conditions, DenseNet-121 achieves a validation accuracy of 94.98% and a test accuracy of 95.32%, whereas ViT-B/16 attains a validation accuracy of 94.71% and a test accuracy of 95.45%, demonstrating clear comparative performance. Additionally, ViT-B/16’s global self-attention mechanism notably reduces misclassification errors, particularly in visually similar gesture classes. Our results highlight the viability and efficiency of transformer-based architectures for accurate, real-time gesture recognition in graphics- intensive HCI applications.