We are trying to improve our YoLo algorithm results of recognizing one class of varying sizes (~ varying distance to the camera). Luckily, the a priori position of the object is known with a certain degree of confidence. Hence we cut out a patch of the image where the object is very likely to be situated in the inference process, e.g. (object: hut, in a 1080p image): Likewise, our model's training set consists of similar-sized patches which we calculate from label data.
In this constrained environment the results are pretty good up a certain size (distance), but the detection mostly fails for small (far away) objects. To accustom our model to small objects, we are considering to pass custom anchor sizes, which we can calculate for every image, dynamically to the training process. However we are unsure as to how to accomplish this without forking and customizing Keras. Any ideas?