Hi Wei Hu,

Thank you so much for testing our code. Your issue just scares the hell out of me... I can reproduce your problem and it is because previously we use the 2x numbers of the ground truth junctions as the number of candidate junctions. This should not be used in the evaluation mode and it is fixed now in the master by changing the number of candidate junctions to a hardcoded constant during testing.

I don't think this change will affect the PR curve and other metrics much since it only increases the number of candidates lines rather than changing their score/ranking. But I will rerun the evaluation code and include a script for the ease of testing on a new image without the hassle of data processing, hopefully in this weekend. Again thank you for finding the problem.

Best,
Yichao Zhou.

Thanks for your reply , I test you updated code , the result seems like have more messy lines than before. Updated code result

Previous results

Maybe the number of candidate junctions is set too large

I will update the thresholding scheme as the current one is arbitrary. (it should be thresholded by the confidence rather than a const count). The visualization threshold might also be affected according to your figures.

I implemented the new thresholding strategy. The performance metric in 4c74116 (using N_gt * 2 junctions) is:

msAP 5/10/15: 58.9 62.9 64.7
APH: 83.0
FH: 81.6

and the performance metric now (thresholding on the junction map) is

msAP 5/10/15: 58.9 62.9 64.7
APH: 82.8
FH: 81.2

I will consider this as fixed.