Qi HE
'grad_clip' setting is also different between these two sources.
看样子要放windows下再重新编译一遍。纯overleaf还不太行
hey, I solve this problem by locate the bug in this line! https://github.com/gaoxiang12/slambook2/blob/f522131fadae7e7ec1a09f18873bc65369cdd073/ch7/orb_self.cpp#L384 the coordinates of point pp and qq can be negative value, so the img.at() crash. I solve...
Hi, I found I didn't add neural refinement during training. Can you also provide a `co3d_nr.yml` for training?
Actually, the ground truth depth is not pixel-wise related to the RGB image. This could be the reason for your poor results.