Open TPGamesNL opened 7 months ago
Side note, but the way the classifier is used is interesting. The output for a certain portion is a probability ([0,1] I think).
They take all the images extracted from the screenshot, and calculate the logo probability. They then take the 3 with the highest probability and use them for reverse image search
Classifier for detecting whether a certain portion of a screenshot is a logo isn't that good.
Decision matrix: (apologies for bad text coloring, vscode is nice like that sometimes)
text version: Confusion matrix of Decision Tree optimized for recall_score on the test data: pred_neg pred_pos neg 35977 43 pos 168 161
it's great at determining that something isn't a logo, but if you feed it a logo it's essentially a 50/50