Hi, sorry to disturb you. But when I reproduce the quantitative performance of my own model only with the standard loss, I find there will be a large difference using evaluation dataset you provided or the origin evaluation dataset from dtu_yao.py
Specifically, when I use the depth maps from dtu_yao.py(using test list) to get point clouds by fusion.py, the number of the points will be dramatically low(about 1 or 2 million), thus make the acc. and comp. too high(about 3.3). But using your evaluation dataset can produce much better result
I'm wondering where the reason lies in.
Hi, sorry to disturb you. But when I reproduce the quantitative performance of my own model only with the standard loss, I find there will be a large difference using evaluation dataset you provided or the origin evaluation dataset from dtu_yao.py Specifically, when I use the depth maps from dtu_yao.py(using test list) to get point clouds by fusion.py, the number of the points will be dramatically low(about 1 or 2 million), thus make the acc. and comp. too high(about 3.3). But using your evaluation dataset can produce much better result I'm wondering where the reason lies in.