Xiaoyang Wu

Hi, the config looks good to me. How about the coverage of the model? The gap between train split and val split. The scale of customized datasets also might be...

Hi, discussion here contains several instruction on accelerating testing process (https://github.com/Pointcept/Pointcept/issues/108#issuecomment-1893709522)

Sorry for the late response. I have been too busy recently, and I have to ensure the progress of my current research project. Actually, Pointcept also supports object classification; you...

What's the meaning of 128 in [B, 128, 512]? First of all, B and N are combined in Pointcept to support dynamic size (here, https://github.com/Pointcept/Pointcept?tab=readme-ov-file#offset) And to reduce it to...

In my opinion, 3D backbones can not be metric by a simulated input with a given shape, as the spatial relation and density can not be simulated correctly. Hence I...

Hi, good suggestion. You can modify InformationWriter here (https://github.com/Pointcept/Pointcept/blob/main/pointcept/engines/hooks/misc.py#L79): add a parameter to control the frequency and enable it in "self.after_iter".

> Hi, thanks for your nice work! In your ptv3 paper, you show an amazing improvement on several datasets, including Waymo. When I run your ptv3 codes on Waymo, I...

Here, use this script to turn our cached prediction during testing to a valid waymo submission: https://github.com/Pointcept/Pointcept/blob/main/tools/create_waymo_semseg_submission.py

Apology for the late response; I didn't notice this issue. And thank @yxchng, for bringing this issue to my notice.

> thank you for your reply! So do you have ever calculated the computation amout of each layer? Never, just benchmark the latency and memory cost. :)