I've wondered about this - can the car "see" a trailer that's flat and has no load on it, as well as it "sees" a full-box type trailer? And I've noticed it doesn't seem to "see" a tow truck very well - had a tow truck pull in front of me recently and if I hadn't taken over it would have rear-ended it for certain. So, I think you're on to something - the car doesn't "see" odd shapes as well as normal shapes.
Neural networks are the state of the art technology for feature and object detection. They consist of a very rudimentary model of how our brains work to achieve the same task - layers of interconnected neurons progressively identify fundamental lines, then features, then shapes, and objects. They are not explicitly programmed - instead, they are trained. The neural network is shown thousands or millions or even billions of examples of objects. The network gradually learns how to recognize the objects that it is trained with.
At a minimum, for AP in the 7.1 release, the distinct objects the processor is trained to recognize include lane markings, speed limit signs, trucks, cars, and motorcycles. Obvious, because this is the feedback we get from the instrument cluster.
The neural network in the current AutoPilot hardware has limited processing ability. It is nowhere near the complexity of the portion of the human brain that we use to drive on the highway - even ignoring street signs, traffic lights, traffic cones, pedestrians, animals, etc etc.
Like you, I've also observed that it doesn't classify certain vehicles very well. There was a picture uploaded in this forum earlier of AP 7.1 identifying a car (sedan, I think) in a drive-through bay as a semi-truck - possibly the tall rectangular outline of the drive-through bay tricked the neural network into identifying the car as a truck. These poor classifications could either be a fundamental limit in the processing ability of the processor hardware or in the breadth training data - or both.
Since identifying the exact type of vehicle isn't critical (versus, say, finding lane markings), I'm guessing Tesla hasn't devoted much time to getting that classification to perform well. Especially for less common vehicles, like the tow trucks or empty trailers that you've noted.