In this talk I will show how “blind” visual inference can be performed by exploiting the internal redundancy inside a single visual datum (whether an image or a video). The strong recurrence of patches inside a single image/video provides a powerful data-specific prior for solving complex tasks in a “blind” manner. The term “blind” here is used with a double meaning: (i) Blind in the sense that we can make sophisticated inferences about things we have never seen before, in a totally unsupervised way, with no prior examples or training data; and (ii) Blind in the sense that we can solve complex Inverse-Problems, even when the forward degradation model is unknown.
I will show the power of this approach through a variety of example problems (as time permits), including:
- "Blind Optics" — recover optical properties of the unknown sensor, or optical properties of the unknown environment. This in turn gives rise to Blind-Deblurrimg, Blind Super-Resolution, and Blind-Dehazing.
- Segmentation of unconstrained videos and images.
- Detection of complex objects and actions (with no prior examples or training).