This is an old revision of the document!
odd things that people do with cameras
or
machine vision
or
understanding images with a computer
as a computational photographer, why should I care about this?
- everything you do with digital photography derives from this field, eg
- digital filtering (from ccd de-bayer algorithms upwards)
- auto focus/exposure/blah = basic algorithmic image understanding
- face finding/smile detection
- a lot of the future will come from this field
- object recognition
- augmented reality
- realtime scene mapping
(will talk of some of this later)
- it is a young and vibrant research area
- lots going on
- much to be discovered
- lots of _hard_ problems to be solved
but not hard to get fun things happening yourself
- as an artist you can cherry pick the easy/fun bits or use good quality prewritten libraries
what computer vision is not:
- anything to do with neuroscience (a lot of people think it is)
- it's all statistics!
example 1 : separate moving and still things in a sequence of images
- frame differencing
- finding overall direction of movement
- this actually works, is robust to lighting changes, camera wobble etc
- results in useable data
- (all eyetoy games use this technique)
why it works - temporal memory is short and well defined
example 2 : seperate background from person in scene
* subtract current image from reference image without persion * works! (for a few minutes)
lighting changes, camera auto settings, passing cars… all conspire against us.
doesn't work
_very hard problem_ seen much time and money sunk into solving this without result
example 3 : faces
faces are great for computational photographers because:
- most images have a few in them
- we all have one
- we are particually attuned to understanding them
- a lot of research has been done on this, a lot we can build on as artists
actually several problems
- recognising what is a face in an image (face finding)
- recognising who is who in an image (face recognition)
- face finding: haar cascades, wavelets, example images
- face recognition: eigenfaces, image space vs face space
eigenfaces
- extracting information on age, gender, ethnicity, expression from a face image
- modifying this information - changing a faces age, gender ….
scary examples
augmented reality
- marker pose estimation
- rendering 3D objects in the real world
- yeah yeah but you need a marker…
SLAM: simultaneous location and mapping
- gradually building a model of the scene from a camera in realtime
- example video
- no markers needed
- lots of cool things will come from this in future
3D cameras
- RGB + depth per pixel
- microsoft project natal
solves a lot of problems, creates some interesting new ones
how?
- stereo cameras
- structured light
- wavefront analysis
where do I get some of this stuff?
- opencv
- artoolkit
- nasa image processing library
- lots of research code is out there and waiting for artistic (mis)use!
- they may actually help you in between publishing papers
tips
- think like a photographer - how can I set the scene to suit what I am trying to do, lights, camera position etc etc are as important as finding the right algorithm.