Incorrect Precision and Recall Metrics

[YoungseokOh]

Hi,

I believe there's an issue with the get_confidence_list() function.

When I used your pre-trained model, I couldn't achieve the same performance metrics as you reported.

I think the else clause should be removed because it calculates values that should not be considered.

The function should only handle cases where the prediction is a true positive and matches well with the ground truth (GT).

If a prediction does not match well with a ground truth, it should not append 0 to the same list (true_positive_list).

Appending 0 incorrectly includes non-matching predictions in the calculation, which creates problems when calculating precision and recall.

I will wait your reply.

Thanks

[Teoge]

For each ground truth, the algorithm attempts to find a matching prediction. If no prediction matches the ground truth (indicating that the model failed to detect the object), the algorithm appends a 0 to the true_positive_list.

When calculating precision and recall, a threshold is used to divide the true_positive_list into true positives and false negatives:

false_negatives = bisect.bisect_left(true_positive_list, thresh)
true_positives = len(true_positive_list) - false_negatives

As long as the threshold is greater than 0, these appended 0s in the true_positive_list will always be treated as false negatives, which is the correct behavior. Ignoring them would result in an incorrect false negative count, leading to underestimation.