I noticed that when calculating precision and recall for validation results, if I merge all images together (ignoring the batch dimension), the accuracy can be very high. But if I calculate precision and recall for each sample in the batch and then sum them, the result is only 88%. I know this is related to the distribution of positive and negative samples. In the field of ice lake detection, which method is more appropriate?

More Wang Qing's questions See All
Similar questions and discussions