Suppose a CNN is trained to detect bounding box of a certain type of object (people/cars/houses/etc.)

If each image in the training set contains just one object (and its corresponding bounding box,) how well can a CNN generalise to pick up all objects if the input for prediction contains multiple objects?

Should the training images be downsampled in order for the CNN to pick out multiple objects in the prediction?

EDIT: I don't have a specific one in mind. I was just curious about the general behaviour.

  • 1
    $\begingroup$ What algorithm are you using? It's dependent on that $\endgroup$ – Abhijit Balaji Mar 28 '18 at 14:51
  • 1
    $\begingroup$ @AbhijitBalaji I don't have a specific one in mind. I was just curious about the general behaviour. Which ones are particularly better at generalising? $\endgroup$ – Kar Mar 28 '18 at 19:05
  • $\begingroup$ Welcome to AI! I added the "software-evaluation" tag, since it sounds like you're looking for performance comparison. (Feel free to add that detail into the body of the question, as not everyone will read the comments.) $\endgroup$ – DukeZhou Mar 28 '18 at 21:18
  • $\begingroup$ "How well" means there should be a measure. Please include this in your question. Otherwise one could simply answer "really well" or "42%" $\endgroup$ – Martin Thoma Sep 27 '18 at 5:57

I suggest you to go through the r-cnn paper or go through a tutorial on it . CNNs transform the image into high dimensional vector in their last layer , in case of classification this vector is sent to a "softmax" layer , in case of bounding box regression , four values :length , breadth , location of one of the points of the bounding box , are regressed from this vector , so if you use a cnn with one regression head you end up with one bounding box irrespective of the training set.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.