I am working in computer vision and have this scenario:
For each frame of a video sequence I have the following:
- Image with a resolution of width * height discrete pixel locations.
- List of "detections" (e.g. from HaarCascadeClassifier) with:
- pixel location (x,y), discrete, where the object (or false alarm) was detected
- size (width, height), discrete, which covers the region where the object is "active"
- List of Ground Truth elements: A list of manually marked objects with:
- pixel location
- size
So I have a 4D object space (2D position and 2D size)...
Now, in theory I have a True Positive sample, if pixel location and size of detection and ground truth elements match, which will never happen in practice. Maybe I can set some kind of "region" around each detection and around each ground truth element and call it a match, if they overlap for some bigger parts...
But what about True Negative Samples? Would this be each location in the 4D object space that isn't touched by either detection elements or ground truth elements?
Does all this depend on the nature of the detector? e.g. the detector works on "windows" within the image. For each window, the detector gives a true/false decision. So I would have to count each of those windows and categorize them to true-positive, false-positive, true-negative and so on? This would make things much easier in theory, but the ground truth will not hit those windows exactly.
I'm asking myself, how true negatives were counted in the original "radar" scenarios of detection theory?!?