Finding key points in a hand image using heatmaps

Abstract

This paper presents an approach to finding key points in the image of a hand using heat maps. For this, a convolutional model of a neural network and a training method on heatmaps were combined. An open bank of palm photos, supplemented by a set of own images and synthetic data, was used as a dataset. Direct and inverse transformations of two-dimensional coordinates on the plane into heat maps with a 2D Gaussian function centered at coordinate points were applied, data was prepared, the model was trained and tested. As a result of this approach, a neural network model able to recognise real world images was obtained.