Efficacy against different angles, lighting, occlusion, etc?

[jonassteinberg1]

Do you have any guess as to what the efficacy will be when images are from different angles and other tough contexts? I recognize this is a question coupled to the data sets themselves and not so much your model, but I noticed while reviewing some of the images in Veriwild that because it's CCTV based many of the images are sequences of vehicles at the same angle and merely "farther down the highway" basically, at least from the few sets that I looked at. So this has me wondering how the model will perform on vehicles coming around a corner? Curious as to youe insight here.