Convergence speed.

[DIVE128]

Hello, we train with a "unification" and “regression” strategy respectively. Abs depth err of "unification" in both avg train and ave test is higher than that of “regression” at the first few epoches. Is it normal to converge slower for "unification" strategy？

[prstrive]

Just looking at the error of the depth map, "unification" does drop slower than “regression”. But from our experience, when the error of depth map decreases to a certain level, the point cloud reconstruction quality does not get better as the depth map error decreases.

[DIVE128]

Thanks a lot. A recent confusion is that "classication" compares prob volumes with supervision signals. From your experience, is it very different in the point cloud reconstruction to use pure prob_volume_pre, softmax (prob_volume_pre ), or sigmoid (prob_volume_pre) when training?

[prstrive]

I think so!

[Sun-Xinnnnn]

What is your Abs depth err value? Mine seems a little big.