search

Karel911 / TRACER

TRACER: Extreme Attention Guided Salient Object Tracing Network (AAAI 2022) implementation in PyTorch
Apache License 2.0
195 stars 41 forks source link

U2Net benchmark comparison #20

Open roimulia2 opened 2 years ago

roimulia2 commented 2 years ago

Hey! Loved the paper.

Have you tried doing a comparison to U2Net? As for now, it's considered the state-of-the-art in SOD. I was wondering if you tested against it.

Karel911 commented 2 years ago

Hi,

Actually, we did not compare TRACER with the U2Net in our research. Based on the reported results of U2Net on benchmark datasets, our TRACER outperformed U2Net for the five benchmark dataset (i.e., DUT-TE, DUT-O, HKU-IS, ECSSD, and PASCAL-S). If you want to know the FPS difference of each model, I recommend measuring FPS under your environment.

dem-esgal commented 2 years ago

Hi, I will share my opinion as well. As I saw during my own "in the wild" testing TRACER has some problems with hairs masking. Integrated metrics such as F-measure or MAE cannot reflect this. I'm sharing non-sigmoid final map boys

roimulia2 commented 2 years ago

Hey @dem-esgal, as for overall "objects", do you think it performs better? (if fine details aren't required)