Machine Learning and Friends Lunch
Sign Detection in Natural Images with Conditional Random Fields
Jerod Weinman
UMass
Abstract
We are building a wearable system to help the visually impaired navigate in a highly visual world by recognizing a wide variety of signs. However, before an object can be recognized or interpreted, its presence must first be detected. Toward this end, we present successful results detecting signs in unconstrained images. Context is an extremely powerful cue for identifying image content, but it is often overlooked or underutilized. Our solution to the detection problem utilizes a probabilistic model that allows more use of image context than previous methods.