Identifying Critical Regions for Motion Planning using Auto-Generated Saliency Labels with Convolutional Neural Networks