Cool! Do I understand correctly, that the splitting into the part where the drone was doing fine and the part where the drone is crashing (i.e. the annotation of the dataset) was still done manually?

A similar approach using unsupervised learning would be even cooler...