I've been working on exploring the impact of blur on convolutional networks with Ayan Chakrabarti and Greg Shakhnarovich at TTIC.
We find that the accuracy of pre-trained classification and semantic segmentation networks drops when they are evaluated on blurred versions of the test set. We explored various fine-tuning strategies, augmenting the dataset with blurred images (with increasing levels of blur, e.g. defocus blur kernel with radius r=2,4,6,8) and find that we can substantially improve accuracy on blurred images without having a large effect on the original sharp images.
I am a contributor to the Zoomout semantic segmentation project with Reza Mostajabi at TTI-Chicago. Zoomout (also called "hypercolumns") is a technique for creating a model for pixelwise labeling through the use of a deep pretrained image classifier.