Microsoft AI for Earth and DrivenData just launched a new machine learning competition on Snapshot Serengeti data:
This is different from previous competitions, where competitors had access to all the data, which inevitably leads to some overfitting on the test/scoring data (even though you don't have labels for that data). We've done a bunch of infrastructure work so that competitor code will run on unseen test data (Snapshot Serengeti data that has not yet been released publicly), which we hope will lead to more generalizable models. More importantly, we hope this starts a trend in the way competitions get run in this area, so we can more quickly "competition-source" real working models for camera trap projects.
Spread the word... looking forward to your submissions!