As a fast.ai starter project I regressed a little model on movie stills from various eras to see if I could predict a year of release given an unseen image from a film. It works reasonably well! (feel free to try https://97f9ebe662372966fa.gradio.live/)
I've been investigating what exactly the model does to accomplish this. I tried Grad-CAM, which suggested the model is not looking at higher-level features of the image like faces, etc. -- that's about what I expected.
Presumably it's instead picking up artifacts from the filmmaking process that reliably vary across time like film grain. But I've struggled to visualize what exactly those are.
I've also looked at using a deconvnet like in this paper (https://arxiv.org/pdf/1311.2901) that's referenced in fast.ai -- but the results from that aren't easy to interpret:
My last attempt was to optimize random noise toward the min/max years in the dataset, but the results from that are basically indistinguishable.
Any ideas for a next attempt?
