Computer Vision & Biomimetic Object Recognition Bruce A. Draper Department of Computer Science January 28, 2008
Background : Computer Vision The computer vision community specializes in the interpretation of image data – 3D reconstruction Stereo analysis (up to N cameras) Motion analysis – Includes image stabilization, image mosaicing, control Mapping & Measurement – Object recognition Model based Knowledge based Learned (supervised or unsupervised) Traditionally funded by the military, but the domain of applications is expanding
Computer Vision Resources CVPR & related conferences since 1983 (PRIP ) – Hosted 1999 CVPR in Ft. Collins – ICCV, ECCV, ACCV, ICPR, ICVS, … Technical Committee of the IEEE (PAMI) Journals – IEEE Trans. On Pattern Analysis and Machine Intelligence (PAMI) – Computer Vision and Image Understanding (CVIU) – International Journal of Computer Vision (IJCV) Machine Vision and Applications (MVA) IEEE Trans. on Image Processing (TIP) Pattern Recognition On-line tools and resources – CVOnline (web site resource) – OpenCV (open library of computer vision algorithms)
Background : Personal Object recognition – Knowledge-based & learned Applications – Face recognition Evaluation of face recognition algorithms & covariates – With R. Beveridge (CS), G. Givens (Stats) Modeling faces as hihg dimensional manifolds – With M. Kirby (Math), C. Peterson (Math), R. Beveridge (CS) – Landmark recognition for self-driving cars Visual where am I? – Automatic population of geospatial data bases Build semantic & temporal maps from satellite images – Biologically-inspired Cognitive Architectures (DARPA BICA) With S. Kosslyn (Harvard) – Counting nesting seagulls on islands off the coast of Maine
What is this? Dirty little secret: computer vision systems can’t do this yet (not in general) Well, there’s a truck, driving over some rocks, with mountains in the background
My goal Learn to recognize objects by mimicing human vision – At the level of regional functional anatomy – End-to-end systems that work! Three examples of how human vision influences design: 1. Selective attention 2. Familiarity detection 3. Goal-directed object detection
Selective Attention Human vision is selective – Overt attention : eye & head movements – Covert attention : internal data selection
Familiarity vs Recognition People recognize whether an image is familiar before they recognize what it is So we show our system (SeeAsYou) a series of images…
Familiarity vs Recognition (II) Then we give it new images, and ask it to retrieve “similar” images from the data set Novel ImageRetrieved Image
More examples
Next… recognition Did we recognize the leopard on the previous slide? – No, the answer was an image, not symbolic Did we match the leopard image? – Depends: we matched it to a cheetah – If the goal was to match spotted cats (or wildlife, or …), we got it right – If the goal was to find leopards, then no. Current research : top-down verification of specific goals based on evidential reasoning
Looking for new applications Image inspection tasks currently done by humans – Rule of thumb : if people can’t do it, neither can our system Object recognition – Not just measurement Lots of data, limited training labels
Thank You Questions?