Another suggestion from Joe was to implement a Siamese Net to incorporate more context surrounding the features. For each sampled point, two crops are made - one with the standard depth proportioned bounding box, and one that is, say, 4x the size, downsampled to the same resolution. Then, two parallel nets are trained, one on the large and one on the small crops, and only the last layer or few layers computes a single label output for both images.
Another suggestion from Joe was to implement a Siamese Net to incorporate more context surrounding the features. For each sampled point, two crops are made - one with the standard depth proportioned bounding box, and one that is, say, 4x the size, downsampled to the same resolution. Then, two parallel nets are trained, one on the large and one on the small crops, and only the last layer or few layers computes a single label output for both images.