How to determine the Threshold of Detector
weiweiBC opened this issue · comments
The number of samples in the results of my detector is much larger than the number of pictures. How do you set the threshold of the detector? So that neither sample loss nor redundant samples can be detected.
Score thresh is 0.05 and NMS is 0.5 to balance obtaining high recall with not having too many low precision detections that will slow down inference post processing steps.
When I set the score threshold to 0.6, the total number of samples detected was equal to the number of labels you gave (such as 30000 + pictures get 50000 + labels). Is my detector not well trained?
It is normal that you detect more instances than the exact ground-truth bounding boxes. That's why we use evaluation metric map for instance detection.
When your score threshold is set to 0.05, will you detect a lot more samples than gt? Another question is, in your competition, if the sample detected is not in gt, will it affect the score of other samples in this picture?
Yes, we will detect a lot more samples than gt. You can refer to 'MAP', the evaluation metric for instance detection. To calculate the AP, for a specific class, the precision-recall curve is computed from the model’s detection output, by varying the model score threshold that determines what is counted as a model-predicted positive detection of the class. In this way, the influence of score thresh and the number of items you submitted is eliminated.