Performance when evaluating custom data about cam2bev

[lreiher]

Thanks for your interest!

Here are a couple of things we have noticed:

• The validation result you have attached doesn't even show the ego vehicle, which is a little surprising, since it's included in all training samples at exactly the same location. You can also see it being predicted in the results shown in our paper. Perhaps something is wrong with the training of the model already.
• Some problems due to the domain shift are probably expected, as you also pointed out. Not only are you using real-world data, but also fisheye cameras and a truck. The truck perspective for example allows to see over cars, resulting in far less distortion in the IPM image.

To further analyze this, could you please post the config file that you were using to train and predict?