search

xuebinqin / DIS

This is the repo for our new project Highly Accurate Dichotomous Image Segmentation
Apache License 2.0
2.27k stars 270 forks source link

Difference between isnet.pth and isnet-general-use.pth #84

Open github-luffy opened 1 year ago

github-luffy commented 1 year ago

Thanks for your great work.

what are the training differences between isnet.pth and isnet-general-use.pth, as isnet-general-use works much better than isnet.pth.

przemb commented 1 year ago

Hi, this information is in README, but I was also initially confused, so I decided to sum it up:

@xuebinqin, please correct me if I am wrong... isnet.pth is a file which was trained only on DIS1 and these weights were mainly shared for reproducibility and benchmarking reasons. ("academic version") isnet-general-use.pth was trained on extended version of DIS1_EXT, which is not available, but it is not the DIS2. DIS1 "includes very few animal, human, cars, etc. So it may not work well on these targets".

Since I calculated HSE metric, I am sharing it to save your time.

Comparison of HCE metrics (smaller number - better performance)

model isnet-general-use.pth isnet.pth
VD 846.068 1116.368
TE1 102.278 148.456
TE2 254.022 340.754
TE3 538.572 686.918
TE4 2338.3 2889.292

isnet.pth column has exactly the same values as the one in the paper :)