Thanks for sharing this YouTube video, it seems great. I believe it depends on how the confidence is calculated in the library. In our case we are considering the "confidence" as the distance between two histograms (two images). In this way, the shorter the distance more similar the histograms/images are.