Adam Kosiorek

One thing is still unclear for me. Some different objects as parts could be very similar to each other. For example a leg of a mailbox or antenna on top of a car (vertical line). And only the context let us to understand the object exactly. What do you do in such cases?

Good question. That's why inference is hard. In this version, we leave inference to an arbitrary neural net, which can take context into account. We impose object -> parts structure only in generation.

Adam, aweome post! When you write that shift-aug. help translation-equiv. CNNs due to the FC layers - is that really the case when you have average/max pooling before them + made the conv-stride Shift Invariant Again (Zhang 2019) Wouldn't it be due to the sub-sampling instead?

Thanks @ikdeepl. I think that if you use the Zhang 2019 trick and max/average pooling then it should be fine without shift-augmentation. Fully-conv classifiers don't usually have linear layers.