Closed songkq closed 4 years ago
Although the precision and recall on these images without foregrounds are ill defined, we didn’t explicitly exclude them when computing the measures. Their measures should be zeros due to the 1e-8 used in the denominator or precision and recall computation.
On Jul 26, 2020, at 12:19 AM, ipScore notifications@github.com wrote:
Greeting! If there are some test samples without a salient object in the image, does it need to contain these samples when computing the PR-curve? Since the precision and recall keep zero with any threshold.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/NathanUA/Binary-Segmentation-Evaluation-Tool/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGOROBMPUCZT3YV35MGFLR5PDIZANCNFSM4PH2OUOA.
@NathanUA Thanks for your reply. I wonder it is reasonable to compute the average PR from the sum of each image? Why not to directly compute the PR from the whole dataset?
Greeting! If there are some test samples without a salient object in the image, does it need to contain these samples when computing the PR-curve? Since the precision and recall keep zero with any threshold.