SaneNet: Training a Fully Convolutional Neural Network Using Synthetic Data for Hand Detection
Amin Dadgar, Guido Brunnett · 2020
We propose a specific formation of synthetic hand images, to train a fully convolutional neural network, for detecting hands in real scenarios. Our methodology aims to achieve this by further exploiting the invariancy concept of these networks. Thus the conventional, and expensive, techniques of domain randomization are not employed to create realistic synthetic images. We train the networks using two types of training images: 1. purely synthetic, and 2. with a combination of a few (100) real images. The results suggest our simplistic and sorely cheap approach of generating synthetic hand could successfully be utilized to detect hands in challenging scenarios.