Are there any existing methods/models describing the probability of an object being detected by a computer vision algorithm given it is seen $n$ times at similar angles and orientations? I know that an autonomous car may, for example, have trouble recognizing a stop sign and as a result the image recognition system being used may have a bounded box around the stop sign continuously appearing and disappearing signaling that the object recognition algorithm is only detecting the stop sign a certain amount of the time it sees it. I would like to understand this phenomenon in a more general sense.

More precisely put, I would like to model that: given an object is detected with probability $p$ if it is seen 'once' (AKA during a moment in time), then what is the probability of it being detected if it is seen for a second time, a third time, etc... Each 'moment in time' may be characterized by a single video frame/image or some other unit of time. Each time the image is seen it may be in a similar but not identical orientation.

  • $\begingroup$ the missing link here, i think, is the identification of the object seen multiple times as the same object in time, instead of as a different (instance of an) object. As a sidenote camshift algorithm can track an object once it is detected (even if it is hidden in some frames) $\endgroup$
    – Nikos M.
    Apr 1, 2021 at 18:52

1 Answer 1


An approach that is not specific to the image domain is to use a probabilistic data structure like a Count-Min Sketch. A Count-Min Sketch data structure can accumulate information to estimate the observed frequency of an input value based on the past set of input values by using multiple hashing functions over the input.

  • $\begingroup$ thanks. I'll think about using this. i'll leave the question as unanswered for now in case someone else offers something better. $\endgroup$ Apr 2, 2021 at 15:52
  • $\begingroup$ can a count-min sketch estimate output values for inputs it hasn't seen before? $\endgroup$ Apr 2, 2021 at 15:54
  • $\begingroup$ The exact behavior of count-min sketch in your application will depend a bit on the hashing functions that are used to process each input image. For each input, it will estimate the observed frequency that input (which for unseen inputs should be a past observation count of zero or possibly greater than zero). The theoretical guarantee is for the exact same input value, the count-min sketch will never underestimate the true observation count. The hashing functions you pick will determine how well this method can generalize over similar (but not identical) images of the same object. $\endgroup$
    – grov
    Apr 3, 2021 at 0:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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