1
$\begingroup$

Object detection is defined as the problem in which a model needs to figure out the bounding boxes and the class for each object. A lot of ML solutions for object detection base around having "two passes" - one for creating the bounding box of the region and another for classifying it.

I was wondering if there is a name for a subset of this problem where $n_{classes} = 1$. I feel like there is an interesting opportunity here as the whole classification part of the model can (basically) be ignored. Obviously, I can just train a typical object detection model with one class, but was just interested to see if there are any more specialized methods.

$\endgroup$
1
$\begingroup$

If you are talking about "two-stage" obejct detectors like Faster R-CNN, note that the second phase is not only for classification, but to obtain more accurate results (https://stackoverflow.com/a/61965140/4762996).

in addition, I guess training a detector with many classes acts like a regularizer, which results in much better accuracies.

The only benefit of explicitly training with one class could be the reduzed model size (and a corresponding speedup).

Note also that there are one stage object detectors (CornerNet: https://arxiv.org/abs/1808.01244, YOLOv3: https://pjreddie.com/media/files/papers/YOLOv3.pdf, DETR: https://arxiv.org/abs/2005.12872, and many more). I just wanted to stress that just because you only have one class, it does not make sense to take a two-stage detection architecture and skip the second stage.

Finally, have a look at: https://stats.stackexchange.com/q/351079

An interesting direction would be to incorporate one-class classification approaches, e.g. as mentioned here: https://stackoverflow.com/a/61965358/4762996.

$\endgroup$
2
  • $\begingroup$ Makes sense, I feel like there is room for improvement by doing it "in one pass". Perhaps this is an area to explore! $\endgroup$
    – agupta231
    Jun 6 '20 at 11:49
  • $\begingroup$ Note then that there are also one-stage object detectors. $\endgroup$ Jun 6 '20 at 11:50

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.