I was wondering what’s the reason that in most routs we can see the car’s hood in the image (front camera), whereas in some routs (Zurich, Zurich_2, Zurich_3) we can’t. I was looking at images in the go_pro_4 subdirectories, episode 0.
All images were recorded with the same rig, however this rig does have camera pitch as a single degree of freedom. Translation and yaw/roll will always stay the same, however, pitch can vary, this is due to the GoPro mount design. I agree that this is not ideal but unfortunately this cannot be changed.
As this dataset should mainly be used for imitation learning from general image composition, as in, the network should learn to understand scenes rather than geometric cues, one can think of variable pitch as an additional image augmentation. One should probably perform image augmentation such as random-cropping etc. anyway.
As a final note, the roof of the car should never be visible but it may be the case that the hood will be for certain routes.
