Temel Kavramlar
Efficient personalization of gaze estimation using meta prompt at test time.
Özet
The article introduces a method for efficient and accurate personalization of gaze estimation without labeled data. It leverages prompt tuning techniques inspired by natural language processing research to update a small group of parameters, known as the "prompt," during personalization. The proposed method aligns unsupervised loss with gaze error through meta-learning, ensuring effective adaptation even with minimal data. Extensive experiments demonstrate significant improvements in accuracy and adaptation speed compared to state-of-the-art methods.
İstatistikler
The proposed method can be 10 times faster in terms of adaptation speed than compared methods.
The prompt can contain less than 1% of a ResNet-18's parameters.
Alıntılar
"Our experiments show high efficiency of the prompt tuning approach."
"Our method significantly outperforms existing methods in terms of adaptation speed and accuracy."