I have a question about the difference between general object detectors and specific object detectors.

By specific object detectors, I'm referring to classifiers/object recognizers that are built to recognize a single object, eg. a person, or a vehicle.

By general object detectors, I am referring to algorithms which are trained on a number of different objects (and are therefore able to differentiate between a car, house, person, and banana). Importantly, I am referring here to algorithms that are inherently designed to accommodate many object classes, not merely collections of several specific object detectors put together.

I have an assumption, which is that: Specific object detectors, which are used to recognize a specific object such as a vehicle, are generally more accurate/precise, than the results to be expected from a general object detector which is being used to recognize that same vehicle, given the same amount of space and time.. The reason is that given limitations on space and time, there is only so much that an algorithm can capture. So using a specific object detector for vehicles would be more accurate since the time and space used by the algorithm's training and testing is specialized on vehicles only and is better at this task.... rather than using a general object detector for recognizing those same vehicles.

Put briefly:

"With an increase in the number of objects to recognize, the complexity of models increases, and in accommodating for that, the reduction in accuracy of detection for each individual object also increases."

I've looked for evidence of this assumption but haven't found any yet. And is this assumption true? Is there any evidence to back up whether or not it's true?

  • $\begingroup$ My suggestion would be that you start by going through the published literature on object detectors, read the published papers, and see if you find evidence for or against your hypothesis. Do you know how to do a literature search? $\endgroup$ – D.W. May 13 '14 at 17:39
  • $\begingroup$ Yes, I've been reading a lot of papers on object classification, especially using cascaded boosting. I have come across evidence that shared learning can sometimes cause a reduction in intra-class prediction accuracy. But I haven't found evidence about the more general field of object recognition. $\endgroup$ – user961627 May 13 '14 at 17:52

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.