I am using computer-vision for detecting objects on the measuring table. In my tests, I used two techniques/frameworks: semantic segmentation (framework TensorFlow) and object recognition (framework Halcon). Both techniques lead to good results, and in this case, every new object needs to be labeled and trained.
- Could this process be simplified, i.e. could the machine be taught how the measuring table (background) looks like and not how objects look like?
It seems that this solution would have many advantages, it would be enough to teach the machine what the measuring table looks like only once. In this case, it would not be necessary to teach the machine what the objects look like (the number of these objects can theoretically be very large). The problem I see here is a small number of different/suitable measuring table images that I can provide. This table does not change at all and theoretically, only one image is available.