Machine vision...
RinksCustoms
Posts: 531
Just a quick obsrevation, while staring at the graphics demo, particulary the rotating stars - upper right, i've noticed an optical illusion (a counter rotating oval inside the stars), and it led me to the observation that humans are great at recognising curves, silouets, colors, and shapes, simultaineuosly all while doing this in stereo. I also noticed that trying to concentrate on the points at the inside of the rotating stars, i found this task much more difficult, and so a theory is born..
Computers have a difficult time deciphering curves, colors, silouets, shadows, ect., in color all simultaineously & in stereo, you can imagine the power needed to do all this at even a modest frame rate. Object recognition is a whole nother can-o-worms, computers have a long way to go to be able to decipher say a human from the room. We as humans proccess visual information at an amazing rate, we already have numerous images stored of the room, and with trained stereo vision over years of expieriencing the world around us gain the ability to separate the book on the table form the table and the room.
Perhaps it's possible to filter a b/w image, train the program to "look" for reference points (ie. a %00 to %01/%10 transition) store it into an array, compare that information with the table holding the filtered image of the room and the difference of the two "images" should yield your object. Of course this brings the problem full circle in that humans can recognise the same object at different angles, colors, textures (luminance vs reflectivity), and training a computer to do this would be very difficult. I would imagine this could be done by storing a 2-axis, 360 degree stereo image file of every object encountered, even a 4 bit b/w image of one object,
2 cameras(2 axis (frame size*color bitscale(360/degrees of res)))
would be a pretty big file. A way of compressing this filesize would be to reduce the resolution of degrees the camera does a snapshot of the image on each axis and teach the software to "guess" at the object later, but with reduced degrees of resolution on each axis and color bitscale, multplied by the size of each frame captured, your "guessing" error increases exponentially, maybe to the point of the whole program being useless. Having the user spend more time telling the computer what a particullar object is rather than having the computer correctly guess what the object in question is.
and thus the realization sets in that without a good object recognition ability, computers will continue to be "blind" being able to only tell colors and basic movement. But there may be a different way to get computers to see. One method could be separating colors from the background by bit transitions (comparing two or more images stored and "looking" for linear movement), and calculate the velocitiy,acceleration, and direction of the "virtual" objects, it isn't foolproof & would probably be full of false triggering/alarms, but would greatly reduce the amount of memory needed.
It boggles the mind reading this and wondering just how much storage and proccessing speed and power the human mind really has.
The application of stereo video (3D) imaging, ie. capturing and displaying is quite possible. Phil Pi has a thread some 103 pages long on the subject, i have not read the full thread due to other things limiting my time here. That thread is called the "PROPcam".
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Definetly a E3 (Electronics Engineer Extrodinare!)
"I laugh in the face of imposible,... not because i know it all, ... but because I don't know well enough!"
Computers have a difficult time deciphering curves, colors, silouets, shadows, ect., in color all simultaineously & in stereo, you can imagine the power needed to do all this at even a modest frame rate. Object recognition is a whole nother can-o-worms, computers have a long way to go to be able to decipher say a human from the room. We as humans proccess visual information at an amazing rate, we already have numerous images stored of the room, and with trained stereo vision over years of expieriencing the world around us gain the ability to separate the book on the table form the table and the room.
Perhaps it's possible to filter a b/w image, train the program to "look" for reference points (ie. a %00 to %01/%10 transition) store it into an array, compare that information with the table holding the filtered image of the room and the difference of the two "images" should yield your object. Of course this brings the problem full circle in that humans can recognise the same object at different angles, colors, textures (luminance vs reflectivity), and training a computer to do this would be very difficult. I would imagine this could be done by storing a 2-axis, 360 degree stereo image file of every object encountered, even a 4 bit b/w image of one object,
2 cameras(2 axis (frame size*color bitscale(360/degrees of res)))
would be a pretty big file. A way of compressing this filesize would be to reduce the resolution of degrees the camera does a snapshot of the image on each axis and teach the software to "guess" at the object later, but with reduced degrees of resolution on each axis and color bitscale, multplied by the size of each frame captured, your "guessing" error increases exponentially, maybe to the point of the whole program being useless. Having the user spend more time telling the computer what a particullar object is rather than having the computer correctly guess what the object in question is.
and thus the realization sets in that without a good object recognition ability, computers will continue to be "blind" being able to only tell colors and basic movement. But there may be a different way to get computers to see. One method could be separating colors from the background by bit transitions (comparing two or more images stored and "looking" for linear movement), and calculate the velocitiy,acceleration, and direction of the "virtual" objects, it isn't foolproof & would probably be full of false triggering/alarms, but would greatly reduce the amount of memory needed.
It boggles the mind reading this and wondering just how much storage and proccessing speed and power the human mind really has.
The application of stereo video (3D) imaging, ie. capturing and displaying is quite possible. Phil Pi has a thread some 103 pages long on the subject, i have not read the full thread due to other things limiting my time here. That thread is called the "PROPcam".
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Definetly a E3 (Electronics Engineer Extrodinare!)
"I laugh in the face of imposible,... not because i know it all, ... but because I don't know well enough!"
Comments