Digging about for current knowledge regarding image annotation mechanisms, I Googled across Jim Gemmell's publications regarding MyLifeBits. The MyLifeBits project is essentially a bid to construct Vannevar Bush's Memex. Run out of MicroSoft's Bay Area Research Center, there's a big emphasis on RDBMS technology. Interestingly, I think Gemmell eschews embedding annotations in the actual media.
For some gut reason, this model gives me the willies. If you embed the annotations in the image (easier said than done) if you have the image you have the metadata, and there's no single point of failure. Also, the data is separated from the name of the object. However, doing queries is way easier, and that actually may be a big enough win.
Unfortunately, the relative lameness of embedded image annotation mechanisms is driving me to this mode of operation anyway.