Annotating Activity Data with Transformer-based Deep Learning Model
Parsa Salehi, Sook-Ling Chua, Lee Kien Foo · 2024
There are many activity recognition systems based on supervised learning that have been proposed over the years. One problem with supervised learning is that it requires sufficient number of labelled data for training. The majority of the labelling tasks are done manually by the user themselves. This process is rather time-consuming and tedious. Although there are studies that attempt to use active learning-based methods to assist in the annotation process, these methods still require some amount of effort from the user and are impractical, especially when implementing a home for the elderly. In this paper, we propose an automated labelling approach using a transformer-based deep learning model to label the daily activities. Our method leverages on spatio-temporal information for class annotation. We evaluated our method on publicly available datasets.