I am trying to train a YOLO (v3) network on a set of images, but I am faced with a problem when preparing the training set.

Sometimes, an image contains my object of interest, but only half of the object is visible in the image (e.g. half a cat, with the rest out of frame). How do I deal with that?

Should I:

  • Still including the bounding box of the partial object in the training set?
  • Only include bounding boxes for objects which are fully in the frame?
  • Only include bounding boxes of objects whose "fractional area in frame" is above a certain threshold? (e.g. objects that are 90% in frame)

And if I need to include partial objects, should I give the bounding box containing the part of the object visible in the frame? Or should I give the actual bounding box (which would thus go over the edges of the images) of the "true", full object (which I actually have access to)?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.