|
Object Recognition Case Study
Tom Benson [5 posts] |
17 year
|
Hello All,
I'm looking to use Roborealm software to do some object recognition. See sample image below.
The goal is to detect a finite set of laboratory objects that are a placed on a tabletop.
We always know the exact orientation of the camera and the table in relation to each other.
The goal is to identify the various objects (beakers, glasses, tools, jars). The objects can be compared to an image database which has been set up ahead of time (the objects will always be the same, all that changes is where they are sitting). Lighting can be controlled to suit our needs. Stereo is a posibility if appropriate.
Hoping for 80% accuracy of recognition.
Ultimately for a given image or set of images, the goal is to identify the object, mark it at the center of mass, and provide an X,Y coordinates reletive to the image (or X,Y,Z if stereo).
You folks here are robovison experts and I'm strictly a newbie! So any advice most welcome.
Thanks,
Tom B.
|
|
|
|
Anonymous |
17 year
|
I also work on samiler project. So far, I only can find the object.
I hope we can work together.
Calvenn
|
|
|
|
Anonymous |
17 year
|
Tom,
Are the object angles always the same? Ie is the camera placement towards the objects always at the same angle? Can the angle change to view the objects from the top of the table looking down?
The reason for the question is to determine how much 3d recognition is needed. If the camera were at the top of the table looking down most of the recognition becomes a circle analysis of size/color which may make the problem much easier.
If not and the angle stays at a constant angle you will be looking for more cone type shapes and will have to add in object seperation filters as objects will tend to overlap from a side angle (as apposed to a top down angle).
Do you also have a sense of what the shape database will include? Again, curious if the database will contain the images represented at a certain angle, same color and/or same size.
Cups and glasses are tricky to recognize since the usual feature detection used in most other object recognition systems will not work. You are instead working with a color/outline shape analysis which can be difficult if many angles are needed.
STeven.
|
|
|
|
Tom Benson [5 posts] |
17 year
|
Steven,
It is possible to take a top-down snapshot of the workspace, it this would make it substantially easier to recognize. This is a typical lab table surface, which means, the objects are exactly as you might expect...mostly round (from the top) and conical or cylindrical if viewed from side. Bottles and beakers and glasses. Let's assume that most of them are sitting squarely on their bottoms, while a few of them are tumbled and lying on their sides. This is a real-world application...I can't tell the people using this space to "put all the beakers back in the same orientation". The countertop is going to be left the way it is left by typical people...i.e. ... messy!
One other issue is that I need the center of mass marked from both a "side view" and a "top view" if you see what I mean. See drawing below...it shows 2 beakers and a dish. Center of masses are marked with Red crosses.
All of them are circular from a top view and one hopes a system could easily mark the center of the circles from that top view. Great. But they are very different heights in the z axis. The dish is extremely short in the z axis while the beakers are medium short and tall. In all three cases I need the center of mass marked from this side view as well.
Does this make sense? Apologies if I'm using imprecise terms but hopefully the point is getting across :)
Finally, you mentioned overlap. Yes, we should expect plenty of overlap when viewed from any angle. Again, this is a messy real-world application. Lots of objects jumbled together. We just recognize and mark as many as we can, hopefully 80%, maybe more, maybe less, depending on how messy it is.
Tom
|
|
|
|
Anonymous |
17 year
|
Tom,
Good points. Makes sense. Getting the cog on both views is possible. You could use one to help locate the other. If you have a sample image from the top that can help to get things started. Not much sense in getting too far ahead without an actual test image as you're right about the world being very messy ... pixels too.
The theory is that the overhead view should be able to detect all the tops of the objects using the circle detector. When objects that are on their sides they might be able to be recognized by their shape. That's step one. Let's try that on a real image and see if that theory is correct.
STeven.
|
|
This Thread
