Maturana and Scherer (also see  for their follow up work) tackle the problem of safe landing zone recognition using 3d convolutional neural networks. While the system itself merely generalizes convolutional neural networks to 3d volumes, trained on occupancy maps as input, Maturana and Scherer also focus on computing the occupancy maps in an efficient and reasonable way — see the discussion in  for details — and on generating appropriate synthetic dataset to evaluate the system. On a high level (also see my notes on  for details), the 3d convolutional neural nework combines one or two convolutional layers including max pooling and ReLU activations with two fully connected layers. Training is done on augmented training sets including rotated input for rotational invariance.
What is your opinion on the summarized work? Or do you know related work that is of interest? Let me know your thoughts in the comments below or using the following platforms: