Dynamic Gesture Recognition by Using CNNs and Star RGB: a Temporal Information Condensation

TL;DR

This paper introduces a novel method called star RGB that condenses dynamic gesture videos into single RGB images, enabling effective gesture recognition using only color data with high accuracy.

Contribution

The paper presents a new technique for representing spatio-temporal gesture information as a single RGB image, facilitating gesture recognition without additional sensors.

Findings

Achieved 94.58% accuracy on Montalbano dataset, matching state-of-the-art.

Surpassed 98% accuracy, recall, precision, and F1-score on GRIT dataset.

Outperformed existing methods by over 6% on the GRIT dataset.

Abstract

Due to the advance of technologies, machines are increasingly present in people's daily lives. Thus, there has been more and more effort to develop interfaces, such as dynamic gestures, that provide an intuitive way of interaction. Currently, the most common trend is to use multimodal data, as depth and skeleton information, to enable dynamic gesture recognition. However, using only color information would be more interesting, since RGB cameras are usually available in almost every public place, and could be used for gesture recognition without the need of installing other equipment. The main problem with such approach is the difficulty of representing spatio-temporal information using just color. With this in mind, we propose a technique capable of condensing a dynamic gesture, shown in a video, in just one RGB image. We call this technique star RGB. This image is then passed to a classifier formed by two Resnet CNNs, a soft-attention ensemble, and a fully connected layer, which indicates the class of the gesture present in the input video. Experiments were carried out using both Montalbano and GRIT datasets. For Montalbano dataset, the proposed approach achieved an accuracy of 94.58%. Such result reaches the state-of-the-art when considering this dataset and only color information. Regarding the GRIT dataset, our proposal achieves more than 98% of accuracy, recall, precision, and F1-score, outperforming the reference approach by more than 6%.