Different from offical
bluesky314 opened this issue · 0 comments
bluesky314 commented
How is this different from the offical implementation? I thought pointrend will only show significant benefits if the coarse map is at a very low resolution like 20X20 and we need to upsample it 4 or 5 times, which is the case in instance segmentation nets. For semantic segmentation, I am not sure how it would help.