4 ms·
Woah. In an ML class I took last fall we got to design our own classification challenge problems that the entire class then competed on. The one our group made
by rococode 8y ago
Woah. In an ML class I took last fall we got to design our own classification challenge problems that the entire class then competed on.
The one our group made used this exact dataset (the museum website)! We called it VANGOGHORNO haha. We basically pulled the images from the collection and mixed them with non-Van Gogh paintings. It was surprisingly a very learnable dataset, the top two teams got around 97% accuracy and many teams got to 90%. We resized the images and included paintings from vaguely similar styles like impressionism so we were surprised people did that well. I think the top team used some kind of transfer learning to identify entities in the paintings and learned from that.
- ArcticCelt 8y agoIs there some downloadable archive file of all painting in the higher resolution or did you simply roll your own web scraper?
- rococode 8y agoWe scraped the site. The website takes a pagesize parameter that doesn't seem to have a limit so we loaded all the images in one page and just downloaded them all.
- xiao_haozi 8y agoI didn't see any, but you could stitch them together, e.g. https://vangoghmuseum-assetserver.appspot.com/tiles?id=5646620347596800 https://vangoghmuseum-assetserver.appspot.com/tiles?id=56466...
- bnegreve 8y agoIn case you're looking for supervised art datasets for machine learning, I made a little art-movement classifier you may be interested in. Demo here: https://bit.ly/2xClNHz https://bit.ly/2xClNHz data and models are available for download. You can also have a look at art500k which is essentially the same thing. http://smedia.ust.hk/james/projects/deepart_ana.html http://smedia.ust.hk/james/projects/deepart_ana.html Their dataset is also available for download.
- asdsa5325 8y agoSounds like a cool project!
- Someone 8y ago”It was surprisingly a very learnable dataset” It also is a dataset where cheating is quite easy. A google search may quickly turn up items you left out of the training set. Do you know whether any teams did this (could even have happened by accident, if people googled to find different views of paintings, and accidentally included pictures not in the training set)?