Boundary loss different from TMNet?

[aluo-x]

Looking over the code, it seems like the boundary refinement loss is quite different from the loss used in TMNet.

It seems like in Total3DUnderstanding, the boundary loss is simply amplifying the chamfer distance for points that exist on the boundary. While in TMNet, there is a neighbor based "smoothing" loss.

[lz-503]

查看代码，似乎边界精简损失与TMNet中使用的损失完全不同。

似乎在Total3DUnderstanding中，边界损失只是放大了边界上存在的点的倒角距离。在TMNet中，存在基于邻居的“平滑”损耗。

你好 可以加你个微信 一起探讨下这篇文章吗

[aluo-x]

Just to clarify, I consider this a very minor issue that is not relevant to the main point of the paper. Total3DUnderstanding is a great work. The original boundary loss in TMNet is slow, and I think by applying a edge loss you can probably achieve the same goal (but with much smaller computational cost).