I'd like to train an object detection model with Tensorflow using MobileNet V2 SSD and have hierarchical classes. E.g. to have classes "animal" and it's sub classes "horse" and "donkey". I expect such model to also detect mule but then rather than predicting it's 50% horse and 50% donkey, tell it's ~100% animal.
I have used pretrained OpenImages Yolo model and seen it working but i don't know how to prepare such dataset in TFRecord format on my own. I tried duplicating all the animal species annotation boxes with class "animal" but now the model tends to only detect animals and no horses or donkeys.
Maybe the solution is to train without "animal" and do some probabilities merging math after inference? But how?
