Optimal threshold for test set about wsolevaluation

[coallaoh]

Thanks for your continued interest :)

We use the oracle score map threshold in the corresponding evaluation set.
If it is the validation set, we use the validation-optimal threshold; if it is the test set, we use the test-optimal threshold. The threshold is not transferred across the val and test. We do this in order to eliminate the dependence of performance on the choice of threshold; in other words, we treat the choice of threshold as part of the evaluation, instead of as part of the method as done in many prior WSOL works. There have been certain cases where mere the major improvements in localisation are led by better choices of score map thresholds while the paper attributes the improvements to other design choices (e.g. feature erasing thresholds). Our performance metric is measuring the upper bound of the box-based localisation accuracy of the given score maps.

Hope it clarifies your question.

[won-bae]

Thanks for the clarification. Now I got it!