Generating Multiview Hand Gestures with Conditional Adversarial Network

Gesture recognition is one interesting and important topic of research. However, many a times the recognition fails due to self-occlusion in the gesture images. To overcome this problem generally, multiviews are used for hand gesture recognition. Unfortunately, not many multiview gesture datasets are publicly available for conducting research on this topic. A solution to this may be to create multiple views of one particular gesture from one given single view of the same gesture using some gesture synthesis technique. In view of this, we propose a Generative Adversarial Network (GAN) based gesture synthesis model that uses conditional translation. Conditioning the translation model provides appearance information for generating the target view. Various combinations of losses are used as objective functions for training the GAN. The proposed method is experimentally evaluated by comparing the artificially generated gesture images to the original multiview gesture images available in the HGM-4 dataset. The generated images are proved to be highly photorealistic, allowing them to be used reliably in multiview gesture recognition research.