The parts-based models that we applied to human body posture recognition for BASE-IT raised several questions. How could we explicitly determine "parts" and train for those? Was there a way to improve the structural models that combine parts into a full object? We are now investigating methods to answer these questions. We have since deviced algorithms that can handle explicitly specified object parts and that can explicitly learn part relations. Capt. Jones' thesis on AK-47 detection (for forensic data analysis) is a direct outcome of this resarch. Our current experiments at Camp Roberts also utilize these methods, but have the additional functional requirements of rotational invariance, structural models for larger numbers of parts, and fast, embedded execution.

Eventually, we hope that our research will help us inch towards a better understanding of the elusive low-level image representations that have long been suspected to be the first level of procesing in human vision. Some very recent discoveries (Functional connectivity in the retina at the resolution of photoreceptors. Field et al., Nature 467, Oct. 2010) shed additional light on primate vision, our guiding light and vision for computer vision. (Okay, bad joke.)

People: Rob Zaborowski, Justin Jones, Samual Toepke, Simson Garfinkel, Mathias Kölsch