Efficiency Evaluation of Mobile Vision Transformers

Since the adoption of convolutional neural networks for computer vision tasks researchers have strived to adapt them to resource limited hardware. Recently, transformers have shaken up the world of natural language processing and also promise to do so in computer vision. However, their strengths are yet to be determined. In this paper we evaluate the efficiency of the most popular mobile vision transformer models in terms of latency and accuracy on ImageNet-1k and optimise their performance on a real-time facial landmark estimation task. Code and models publicly available at: https://github.com/pcr-upm/icits24_landmarks .