If you can figure out what that attribute signifies based on the rank ordering of the linked sample, I’d be willing to entertain that this is creepy. I can’t make heads nor tails of what that...
If you can figure out what that attribute signifies based on the rank ordering of the linked sample, I’d be willing to entertain that this is creepy. I can’t make heads nor tails of what that attribute is intended to signify, myself. My best guess is that it is extracting a feature that would be used for ML models—likely indicating whether there is some kind of action occurring in the photo. But, it doesn’t seem quite reliable at that, so maybe that’s not what it’s for. I.e., the first and third photos look like action shots, but the second photo does not, despite a ZBEHAVIORALSCORE score of >=0.90. If that’s the case, I don’t think this is creepy, just a case of a bad variable name.
Computational aesthetics is a pretty interesting problem space, but without having benchmarks to validate how well these metrics perform, I’m skeptical. It would take a lot of work to validate...
Computational aesthetics is a pretty interesting problem space, but without having benchmarks to validate how well these metrics perform, I’m skeptical. It would take a lot of work to validate yourself with your own photos—I’m not about to go through my library and manually rank every photo, then see how it compares to Apple’s models. Some of these things don’t need to be as good as a human, though. They just need to be high precision. That is, as long as the highly rated photos look good to most humans, it’s OK if the best-human-rated photo is not the best-machine-rated photo. That is, high precision here is likely OK at the expense of recall.
The attributes in the ZCOMPUTEDASSETATTRIBUTES table is super interesting. Stuff like this hints at what Apple R&D is likely working on and may show up in future products, or what future products are under active development. One can imagine that scene and object classifiers would be essential for augmented-reality products.
Is the ZBEHAVIORALSCORE something having to do with the algorithm or is is as creepy as it sounds?
If you can figure out what that attribute signifies based on the rank ordering of the linked sample, I’d be willing to entertain that this is creepy. I can’t make heads nor tails of what that attribute is intended to signify, myself. My best guess is that it is extracting a feature that would be used for ML models—likely indicating whether there is some kind of action occurring in the photo. But, it doesn’t seem quite reliable at that, so maybe that’s not what it’s for. I.e., the first and third photos look like action shots, but the second photo does not, despite a
ZBEHAVIORALSCOREscore of >=0.90. If that’s the case, I don’t think this is creepy, just a case of a bad variable name.Computational aesthetics is a pretty interesting problem space, but without having benchmarks to validate how well these metrics perform, I’m skeptical. It would take a lot of work to validate yourself with your own photos—I’m not about to go through my library and manually rank every photo, then see how it compares to Apple’s models. Some of these things don’t need to be as good as a human, though. They just need to be high precision. That is, as long as the highly rated photos look good to most humans, it’s OK if the best-human-rated photo is not the best-machine-rated photo. That is, high precision here is likely OK at the expense of recall.
The attributes in the
ZCOMPUTEDASSETATTRIBUTEStable is super interesting. Stuff like this hints at what Apple R&D is likely working on and may show up in future products, or what future products are under active development. One can imagine that scene and object classifiers would be essential for augmented-reality products.