0
$\begingroup$

In object detection, they usually resize by keeping the ratio the same as the original image, which usually names "letterbox" resize.

My question is:

  1. Why we need to do that? As I see with some images too long in vertical or horizontal, we will lose a lot of features in those images.
  2. If it is better than normal resize, why people don't apply it in the classification task?
$\endgroup$
1
$\begingroup$

While many neural network architectures -- including those commonly used in object detection -- can deal with inputs of variable size, it's typically desireable to resize inputs to the same size to improve computational efficiency via "batching", since the workload can be more effectively parallelized.

In object classification, the aspect ratio isn't that important to determining what an object is -- other than some pathological examples (for example, classifying ovals versus circlse). However in object detection, the objective often includes outputting the precise bounding box of the object, and it does matter what the aspect ratio is there -- you could imagine humans always have a width-height ratio of rougly 1:4, and the network might learn this fact as long as the aspect ratio is kept consistent.

$\endgroup$
1
  • $\begingroup$ Thank you for your answer, do you know why they need an anchor box? As I read in the YOLOv2 paper, he just told that use anchor and predict the scale factor is much better than predict directly the bounding box (the model was hard to predict the small object), but didn't tell why $\endgroup$
    – CuCaRot
    Nov 2 '20 at 6:13

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.