Towards Accurate Multiple Human Tracking Using Scalable Distributed Deep Learning

Scalable distributed deep learning is widely studies for open datasets for visual object recognition, i.e., ImageNet. In general, when we apply these scalable techniques to real applications, costly model training to domain-specific datasets is required for accurate recoginition; however, there are a few case studies for distributed deep learning except for ImageNet in terms of scalability, hyper parameter settings, and generalization, etc. This paper demonstrates our early activities on accurate human detection from soccer video images using distributed deep learning, as an instance towards accurate multiple human tracking from application-domain-specific video images.