Xiaoyang Wu

> I didn't do any training and want to jump straight to testing. Hi, this case, I think you can directly run our test script without any modification on config...

Hi, ~~I think for the latest version of Pointcept pointgroup_ops has already become an optional install~~ (Oh, Thanks for your reminder, I just made the installation of pointgroup_ops optional now,...

Hi, for the bug you mentioned, the solution is: `("color")` -> `("color",)` as `("color")` is euqal to `"color"`. Also, I think `intensity` is similar to the concept of "strength" in...

> oint cloud data and visualize result of the segment Hi, In this case, we need to train a new model with only a 3-dimensional feature (xyz?).

Hi, maybe you can check my comment here(https://github.com/Pointcept/Pointcept/issues/108#issuecomment-1893709522) and see whether the strategy here can make you run our official test process.

So, if your point cloud is labelled sparsely, you might lose them during Random GridSampling and SphereCrop. If they are sparse, I think you can modify these sampling strategies, and...

OACNNs are voxel-based methods, yet PTv3 is a serialization-based method, as defined in our paper. As point cloud in PTv3 is structured and ordered.

> > Hi, the results should have the same density as the original point cloud although we adopt two grid sizes. Pointcept will interpolate the downsampled point cloud to the...

Currently, it is. non-trivial to do inference on CPU, as some of our components are built upon CUDA code. Replace SpConv with OctreeConv and detect PTv3 from the Pointcept during...

> > Hi, it might be because the original point cloud is too huge and there are too much of pieces of sub-point cloud need to be predict during testing...