A Survey on: Application of Transformer in Computer Vision

Zhang, Zhenghua; Gong, Zhangjie; Hong, Qingyang

doi:10.12792/icisip2021.006

Cited by 4 publications

(1 citation statement)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…(c) Computational complexity: The Transformer model may require more computational resources and memory when processing longer input sequences [51]. Particularly in the case of continuous writing processes, if the real-time execution of large-scale transformer models is necessary, it may impose increased demands on computing devices.…”

Section: F Methods Limitationsmentioning

confidence: 99%

An End-to-End Air Writing Recognition Method Based on Transformer

Tan,

Tong,

Matsumaru

et al. 2023

IEEE Access

View full text Add to dashboard Cite

The air-writing recognition task entails the computer's ability to directly recognize and interpret user input generated by finger movements in the air. This form of interaction between humans and computers is considered natural, cost-effective, and immersive within the domain of human-computer interaction (HCI). While conventional air-writing recognition has primarily focused on recognizing individual characters, a recent advancement in 2022 introduced the concept of writing in the air (WiTA) to address continuous airwriting tasks. In this context, we assert that the Transformer-based approach can offer improved performance for the WiTA task. To solve the WiTA task, this study formulated an end-to-end air-writing recognition method called TR-AWR, which leverages the Transformer model. Our proposed method adopts a holistic approach by utilizing video frame sequences as input and generating letter sequences as outputs. To enhance the performance of the WiTA task, our method combines the vision transformer model with the traditional transformer model, while introducing data augmentation techniques for the first time. Our approach achieves a character error rate (CER) of 29.86% and a decoding frames per second (D-fps) value of 194.67 fps. Notably, our method outperforms the baseline models in terms of recognition accuracy while maintaining a certain level of real-time performance. The contributions of this paper are as follows: Firstly, this study is the first to incorporate the Transformer method into continuous air-writing recognition research, thereby reducing overall complexity and attaining improved results. Additionally, we adopt an end-to-end approach that streamlines the entire recognition process. Lastly, we propose specific data augmentation guidelines tailored explicitly for the WiTA task. In summary, our study presents a promising direction for effectively addressing the WiTA task and holds potential for further advancements in this domain.INDEX TERMS Air writing recognition, Transformer model, Human-computer interaction (HCI)

show abstract

Section: F Methods Limitationsmentioning

confidence: 99%