how to lower false positive ratio on object detection using negative examples

Question

There is a similar question here but the answer is not so clear;

Basically, I have a model that detects only and only "matchbox". However, it has a high false positive ratio specially confronted with other boxes.

I thought to add other boxes as negative examples; I am using label studio - I am not sure which path to pursue,

1. how to label them ?! (should it be another class which is "no-box" and the coordinate is the entire picture ? or rather

2. no label - meaning there is no coordinate and no additional class. in this case how can I feed it into TF2 pipeline ?

Answer 1

You can create a separate binary classification model that receives the box image and tells apart matchbox vs. not a matchbox.

Answer 2

One option is to create a new class "a-box-but-not-a-matchbox". Label the false positives as that new class and retrain the model.

Another option is to create an anomaly detection system. Learn the distribution of the "matchbox". Any image greater than a distance threshold from that distribution would be labeled as "not matchbox". K-nearest neighbors algorithm (k-NN) is a simple example, there are more complex examples.