This looks like an interesting approach for practical object detection. However I think the generation of part models is going to be the biggest real world constraint. I not sure how effective this approach will be in the long term unless the PR2 is capable of generating the internal models from objects found in the environment as opposed to generating the part models in an automatic photo booth.
Considering the insanity of perpetual copyrights and the modern legal system, I'm not sure that it is possible to develop a system for sharing part models. Without a method for generating models in the field or a system for sharing part models I am not sure how this approach is going to scale up to the consumer level.
If I look at a copyrighted image, my memory isn't legally controlled by the copyright owner, I can use that memory to recognize objects without paying them, but a robot's memory of a copyrighted image may well be. Is the FFT of a trademark controlled by the trademark holder? Is it copyrightable? Robot lawyers may be needed in the near future to resolve this, though the route to personhood may be worked out for AIs, they apparently just need to incorporate.
More information on the ROS object recognition infrastructure can be found here. Though it is appears to be work in progress with limited documentation for generating your own part models and templates.