@WholeMarsBlog Our NNs didn't have any temporal & spatial memory or surround video auto-labeling back then – everything was knitted together in C++ from single frames with manual labeling.

Still much to improve.