|
Object Recognition - Any Idea
David Chik from Japan [31 posts] |
10 year
|
I am looking for a method to distinguish a toy baby as shown in the attached image. Cannot use color since it is the same as human skin color. Cannot use face since it will be confused with other human face. Cannot use 2d shape since it changes with viewpoint (and can RoboRealm consider 3d shape?). Any idea?
Thank you very much.
|
|
|
|
David Chik from Japan [31 posts] |
10 year
|
To write it more clearly, I mean I want to distinguish this toy child from ANY background (e.g. a person brings it around and the robot still keeps track of it). Now I am using fiducial being tagged on the tummy of the toy, but is it possible not to use fiducial? Thank you very much.
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
David,
You'll need many more images to actually know that 'any' background will work. Its not possible to identify this shape on its own (like a fiducial) since its mainly one color with little internal detail. The shape wrt the background might be used but that can only be tested with many examples images. In these types of cases, we typically recommend having a minimum of 100 images and more ideally 500 in which to test against.
Also note that this is typically per object side so if the object can be seen from many different angles as you mention you'll need images for that too. I.e. a 3D object needs to be represented in many 2D images.
It is possible to match a 3D model of the object too but that's not something we do. That technique is present for human based matching (i..e skeleton tracking) but I doubt that has been generalized to the point that you can use it.
To give you an example, the face detection systems these days will process about 1000 images just for a frontal view. Since you only have one exact type of object (i.e. each doll looks like the other) you will not need as many but not by much. To create a recognition system that can recognize an object from many angles is still an active area of research.
I realize that you'd probably have hoped not to put that much time into this but unfortunately this isn't an easy problem regardless of the platform you use. There again, we may also not be aware of some new technology that might make it easier for you ... but we're not aware of what that might be.
The other possibility is to ask why? As in what's the point of this project? Perhaps based on what is needed something less complex can be used to solve the problem. You mentioned the fiducial is not possible. Is that because it will be ripped off, or you cannot add that in some way to the doll? What's the purpose of tracking this in the first place?
STeven.
|
|
This Thread