Slashdot Mirror
Seeing the Forest For the Trees
swframe writes "A new object recognition system developed at MIT and UCLA looks for rudimentary visual features shared by multiple examples of the same object. Then it looks for combinations of those features shared by multiple examples, and combinations of those combinations, and so on, until it has assembled a model of the object that resembles a line drawing. Popular Science has a summary of the research. I've been working on something similar and I think this accomplishment looks very promising."
David Marr proposed the idea of a primal sketch as the first stage of converting the two-dimensional image on the retina to a full understanding of what is being looked at. This work culminated in a paper published in 1980 called "Theory of edge detection."
Marr was a faculty member at MIT, so it is appropriate for this work to have been done there.
For more information, see:
http://en.wikipedia.org/wiki/David_Marr_(neuroscientist)
and
http://www.amazon.com/Vision-Computational-Investigation-Representation-Information/dp/0716715678
-Todd
Omne ignotum pro magnifico.
I'VE been working on a similar project as well!
Maybe if enough of us with the same project interests get together, we can create an accurate summary of the parent!
You see? We could look at each other's projects for combinations of features shared by multiple examples, and combinations of those combinations, and so on!!??
This is amazing!
Honda just gets on with implementing it. Oh, look, it's even got an automobile analogy: Asimo just did a drive-by on your research.
If you were blocking sigs, you wouldn't have to read this.
Even neural nets have to be programmed at some level to exhibit behaviour that the programmers think will allow them to learn the task at hand unless these guys used some sort of genetic algorithm. The article doesn't mention it. Does anyone know?
Also it doesn't explain whether the system just recognises similar pictures to what its seen before - eg this picture looks like object type 123 (which to a human would be a horses rump) or whether it can combine all views of an object and recognise them all as that object , eg this picture looks like a horse. If its the latter how does it do it - does it have to be shown the object from a large number of angles or can it just infer from a couple of angles what the object would be like from many others?
You know, two people or groups can arrive at the same conclusion, because it was obvious in the first place. And why is it so appropriate? What if the work had been done elsewhere, would that be inappropriate or offensive?
I should have been more specific in my first post.
David Marr's vision book (published in 1982 after his early death in 1980) is considered a seminal work in understanding human visual processing.
Marr was trying to describe how humans see. The new work at MIT is trying to allow computers to see. David Marr would be glad to see the developments, whether at MIT or elsewhere.
-Todd
Omne ignotum pro magnifico.
I certainly haven't worked in this area but for years have wondered how people including fairly young children recognize a dachshund, a bulldog and a great dane as dogs and other things as goats, cats, etc. Dogs are amazingly varied in shape and size and color. It seems like a VERY hard problem.
Theres no way just from looking at pictures of dogs that you could tell they're all the same species. There are some characterstics that some breads have in common with others (other than the obvious 4 legs etc) but they don't all overlap. With something like this its more than a simple case of pattern recognition - its aquired knowledge.
Hierarchical models of object recognition are decades old, as are attempts to implement them. This work doesn't yet work better than other engineering solutions, and it isn't obviously any more plausible than other approaches. So, it's a nice start, but it isn't a breakthrough.
No, because it doesn't.
Upper paleolithic european cave art used continuous, flowing lines, created by spit-painting (think prehistoric mouth airbrush), not short, overlapping, straight lines. The system described in TFA produces results that resemble the sort of lame, pseudo-cubist drawing one saw in art schools in the mid 20th c.
Which should have been included into TFA from the start:
http://people.csail.mit.edu/leozhu/paper/RCM10cvpr.pdf
The main achievement claimed is that no image labeling or any additional data like viewport position was needed, the learning process was completely automated.