search

lhc1224 / Cross-View-AG

Official PyTorch Implementation of Learning Affordance Grounding from Exocentric Images, CVPR 2022
MIT License
47 stars 2 forks source link

The difference between result of test and gt #6

Closed ccmCCMfk closed 1 year ago

ccmCCMfk commented 1 year ago

This is my result of test.py: kld = 1.521207194771439 sim = 0.34196439919030536 nss = 0.9306758723494548 but when I try to visualize the affordance region in test.py,it has a big difference contrast to the ground truth region. result: swing_baseball_bat_baseball_bat_000452 png ground truth: swing_baseball_bat_baseball_bat_000452 pnggt result: swing_baseball_bat_baseball_bat_000595 png ground truth: swing_baseball_bat_baseball_bat_000595 pnggt

lhc1224 commented 1 year ago

The task is a weakly supervised setting and does not have accurate pixel-level labelling as supervision, thus some examples show relative location that is not accurate enough.