Ideally pose and lighting wouldn't matter as much as it currently does. I think using a better model to produce feature vectors could achieve this, or perhaps

Ideally pose and lighting wouldn't matter as much as it currently does. I think using a better model to produce feature vectors could achieve this, or perhaps even finetuning the feature model to match human preferences.