1

Is it possible to train an Object Detector (e.g. SSD), to detect when something is not in the image. Imagine an assembly line that transports some objects. Each object needs to have 5 screws. If the Object Detector detects 4 screws, we know that one is missing, hence there is an anomaly.

Actually this is an Anomaly Detection task where there is something else than a screw (e.g. a hole), but unsupervised anomaly detectors are hard to train and not as stable as object detectors.

Is my assumption correct, that even though it is not really an object detection task, one can use such methods?

oezguensi
  • 205
  • 1
  • 8
  • Your assumption is correct. You could use object detection. Perhaps the real question is more around whether object detection is a good way to do it vs your other alternatives. I don't have particular experience with this sort of problem but if you put it in front of me, the first thing I'll try is to train an autoencoder, and then use the embeddings from that in a binary classifier (defect/no-defect) which can be one of: SVM, MLP, use a distance metric with a hard-coded threshold. – Alexander Soare Nov 26 '20 at 19:41

0 Answers0