Closed LinyeLi60 closed 2 years ago
Hi Linye, the visualization.py
is the correct one, and the search_param.py
is actually not used.
Note that the threshold in equation 11 is only used for inference/visualization purposes, which can be found in visualization.py (Line113). For quantitative evaluation, we do not need threshold.
the threshold in equation 11 is only used for inference/visualization purposes
so the threshold in equation 11 will not influnce the evaluation results? In my opinion, the threshold is a very important hyper-parameter because it will decide a sample is 'unknown' or 'known' class, which will influnce the evaluation results greatly. Forgive me for asking such a levelless question.
No, threshold will not be used for evaluation metrics such as AUROC, AUPR, and OSDR, because these metrics calculate the area under a certain curves, which are determined by varying thresholds from 0 to 1.
However, threshold definitely will impact the results using other metrics like micro- or macro-F1, precision, recall, etc. Therefore, to get rid of the impact from threshold values, we encourage the whole community to discard those threshold-dependent metrics for OSR/OOD tasks.
If you still want to get the threshold using training dataset, you could refer to the AFSD/thumos14/threshold.py.
Thank you! My work is about open set object detection, so I need to get the threshold to get better mAP results, is it right?
Right, mAP is also impacted by threshold. But please remember that only training set can be used to get threshold for mAP calculation.
Ok. Thank you for your patient guidance !!!
Closed.
Hi wentao, I can not figure out where is the code to get the unknown threshold of equation 11 in paper. I find visualization.py and search_param.py, so which is the right one ? Thank you!