Please could you point me towards an algorithm that takes a series of object detections performed on a frame by frame basis as input and outputs a filtered series of object detections in which the filter exploits temporal information to enhance the confidence and accuracy of the raw detections.
So for example, I would expect that a series of very low confidence detections of a particular class that forms a coherent trajectory across 10 frames would give an output of increasingly higher confidence over time - if this object is being consistently detected at a confidence of 5% then it is probably really there. Similarly, an object that has been detected in the middle of a single frame at a confidence of 15% but not in any previous frames is probably a false positive and should be filtered from the output or have it's confidence score reduced.
I would expect such an algorithm to retain and consider the past N frames and then perform some sort of multi-hypothesis energy-minimisation process?
