Pose estimation for carvana dataset
I took all the Carvana data (everything they distribute as test and train for their image masking challenge) and adapted it for pose estimation by splitting it as follows:
Training: 3943 examples of each pose
Validation: 1315 examples of each pose
Test: 1314 examples of each pose
This corresponds to a 60/20/20 split between train/validate/test.
There are 16 poses for each car, evenly sampled across 360 degrees of yaw.
Here are some samples