search

CipiOrhei / eecvf

The End-to-End CV Framework (EECVF) goal is to offer a flexible and dynamic tool for researching and testing concepts without the need for the user to handle the interconnections through the system. To better overcome the continuous development of the EECVF it is constructed as a modular and scalable concept. 
MIT License
8 stars 2 forks source link

Can't find Pixel Correspondence Metric #8

Open illided opened 1 year ago

illided commented 1 year ago

Hello

In your articles you always add that it's possible to evaluate edge maps with both FoM and PCM (CPM). I am trying to find benchmark for PCM metric, but the only ones I could find is IoU, FoM and F1. Maybe it was deleted during development? Or it's implemented in another branch/repo?

CipiOrhei commented 1 year ago

Hello, If you look in Benchmarking.run_bsds500_boundary_benchmark you will see that PCM is used. But, be aware that the BSDS benchmarks works only under linux.

CipiOrhei commented 1 year ago

But you made a good point, i will make this more clear in the next update. Thank you for using EECVF and if I can be of assistance please do not hesitate to ask.

illided commented 1 year ago

Found it, thanks! But now I have another problem) How do you calculate PCM? In your implementation I see terms precision and recall, but there is nothing like that in PCM definition. image image image

CipiOrhei commented 1 year ago

Hello, If you look inside https://github.com/CipiOrhei/eecvf/tree/main/Benchmarking/bsds500/bsds/evaluate_boundaries.py you will see that PCM is used to determine if the pixel is a True Positive or True Negative. Having TP, FP, TN, FN we can calculate P, R, F1. Those are generated in notepad file in Logs/Benchmarks. The format should be stated in the file.