1	What’s going on in Computer Vision at the University of Maryland David Jacobs
2	What I’m going to talk about What computer vision is. What vision is. Computer vision. Landscape of research at Maryland. What it’s like to do vision research. Tools used. Research groups. Future careers.
3	Vision ``to know what is where, by looking.’’ (Marr). Where What
4	Why is Vision Interesting? Psychology ~ 50% of cerebral cortex is for vision. Vision is how we experience the world. Engineering Want machines to interact with world. Digital images are everywhere.
5	Vision is inferential: Light
6	Vision is Inferential: Geometry
7	Vision is Inferential
8	Vision is Inferential: Prior Knowledge
9
10	Computer Vision Inference à Computation Building machines that see Modeling biological perception
11	Vision at Maryland Laboratory founded by Azriel Rosenfeld in 1965 One of (maybe the) oldest computer vision labs. Many alumni throughout world. Very big. ~ Four teaching faculty, ~ Nine research faculty, ~40-50 grad students.
12	Research Themes Very diverse. Understanding images of people. Tracking: where are people? In what position? Recognition: Activity; expression; identity. Structure-from-motion. 3D from video. New cameras. Document/video understanding Recognition: lighting, shape.
13	Visual Surveillance-Goals
14	Background Subtraction
15	Tracking and Activity Classification
16
17
18	Reconstruction Video
19
20
21
22	Tracking – Commercial Application
23	Tracking
24	Tracking
25	Tracking
26	Tracking
27	Motion – Commercial Application
28	Recognition - Shading
29
30
31	Classification
32
33
34
35
36	Expression recognition
37	What is it like to do computer vision? A very diverse field Important work comes from people belonging to many different fields: Computer Science; Electrical Engineering; pure math; applied math; physics; neuroscience; psychology. Common language is math Diversity of goals: building useful systems; understanding biological vision; fundamentals of vision; testbed for learning or optimization. Diversity of tools: math; optimization; system building (including real time systems); AI; learning…
38	Modes of Operation: Possible steps in doing a vision thesis Starting points A practical problem: activity recognition; leaf recognition; real-time tracking. An unresolved issue in a fundamental problem: role of occlusions in stereo; effect of motion on pose and shading. A technique: fast multipole methods; Kalman filtering; belief propagation.
39	"First steps" First steps: Implement benchmark algorithms. Find a simple, toy domain to solve mathematically Look for flaws in current algorithms and try to improve them. Influential work: Solve a fundamental math problem, especially one that gives rise to new algorithm. Show importance of technique from other field. Create a new problem. Build impressive system that demonstrates new ideas or potential effectiveness of existing ones.
40	What makes a good vision researcher? Core competency in many areas: programming, math, knowledge of other fields. But can excel in many ways: System building Math/Algorithms Vision and …(graphics, hci, psychology, learning).
41	Where is Computer Vision Going? More Data, Faster Machines => More Interaction with Other Fields. Fundamental Problems Remain
42	Optimization Partly push of video meant bigger optimization problems. Early 90s, SVD, Gradient Descent, Filtering. More recently, Graph Algorithms, Particle Filtering, Mean-shift, Multi-grid, Fast Multipole Methods …. Techniques from theoretical CS, applied math, physics, learning, ….
43
44
45
46	Learning Recognition using supervised learning. Given examples of an object Use classifiers: eg, SVMs, Winnow, Boosting. Grouping using unsupervised learning. Eg., E-M Probabilistic Modeling Eg., Graphical models, texture, ….
47
48
49	Graphics Common interest in modeling reflection, light, 3D shape. Image-Based Rendering.
50	"input" input depth image novel view
51
52
53
54	Biomedical Engineering Segmentation Identify organs to measure them. Find tumors. Tracking Is a heart beating properly? Is there dead tissue? Registration/Matching. Positions of Tumors in Surgery.
55
56
57
58
59
60	Big Data Sets Image Data Bases Kodak, commercial data bases w/ tens of thousands of images. Internet, with millions? Satellite imagery (Petabytes). Monitoring effects of climate change. Custom Data Sets
61