Hi Liu, Do you and your team have any plan to open source the code on S3DIS (ei. data pre process)? Or could you give some idea about how to reproduce it? Many thanks.

Thus can you please give your official testing code?

I have used the test.py to test your pre-trained model (pointgroup_run1_scannet-000001250.pth). I have used intersectionAndUnion function that I want to calculate the IoU for each sample. It seems the performance it's very poor, like only 53.36 mIoU, while you said it should be 68.96. Thus can you please give your official testing code? Many thanks.

Thanks for your great work. I met some problems at the step of training (the installation and compilation steps have been completed without problems). My environment is set the same as README, and spconv has been compiled and installed.

The code seems to require two .so files, "libspconv.so" and "spconv_utils.so". But I only found "libspconv_nms.so.1.0" and "spconv_utils.cpython-37m-x86_64-linux-gnu.so.1.0" at "One-Thing-One-Click/3D-U-Net/spconv". So when I run "python train.py --config config/pointgroup_run1_scannet.yaml", it raised two errors as below:

OSError: /home/xxx/One-Thing-One-Click/3D-U-Net/spconv/libspconv.so: cannot open shared object file: No such file or directory at File "train.py", line 141, in from model.pointgroup.pointgroup import PointGroup as Network File "/home/xxx/One-Thing-One-Click/3D-U-Net/model/pointgroup/pointgroup.py", line 9, in import spconv File "/home/xxx/One-Thing-One-Click/3D-U-Net/spconv/init.py", line 27, in torch.ops.load_library(_LIB_PATH)

and,

ImportError: cannot import name 'spconv_utils' from 'spconv' (/home/xxx/One-Thing-One-Click/3D-U-Net/spconv/init.py) at File "train.py", line 141, in from model.pointgroup.pointgroup import PointGroup as Network File "/home/xxx/One-Thing-One-Click/3D-U-Net/model/pointgroup/pointgroup.py", line 9, in import spconv File "/home/xxx/One-Thing-One-Click/3D-U-Net/spconv/init.py", line 19, in from spconv import utils File "/home/xxx/One-Thing-One-Click/3D-U-Net/spconv/utils/init.py", line 16, in from spconv import spconv_utils

Could you give some advice on how to fix these errors? Thanks for your time.

Hi, thanks for releasing the code. I have q question regarding the superpoint of s3dis you used in the paper. I found that the semantic label inside the superpoint is not same. In this way, how to calculate the label for the superpoint?

Hello, I followed your instructions and got everything (environment) ready to run the code. However, I found it difficult to run the self training part.

CUDA_VISIBLE_DEVICES=0 python train.py --config config/pointgroup_run1_scannet.yaml 

It shows ImportError: cannot import name 'spconv_utils' from 'spconv' (/.../One-Thing-One-Click/3D-U-Net/spconv/__init__.py) I managed to fix spconv by reinstalling it, and it can be imported smoothly in other folders.

Kindly notice that there is a folder /.../One-Thing-One-Click/3D-U-Net/spconv called spconv, and python will automatically refer to this folder while doing like import spconv. But the folder is apparently not the right path that can start up spconv, the right path is /.../One-Thing-One-Click/relations/lib/spconv right? As we mentioned in the installation part.

It's not a proper way to understand what this folder /.../One-Thing-One-Click/3D-U-Net/spconv is used to do by reading EVERY FILE in it.

So I got a little bit confused. Can you please give some instructions on this issue? ^w^

Hi, I was checking the code. I found out that you choose the random segment of each instance to be annotated(https://github.com/liuzhengzhe/One-Thing-One-Click/blob/6b452493727e8fe813c85a97e6e1876767c4c053/data/prepare_data_otoc.py#L95). And at the end of this function you set all points in this segment to be annotated. For example, at scene0000_00, there are 67 instances in total, from my perspective, only 67 points should be annotated. I calculated len(np.where(sem_labels!=-100)[0]) for this scan and I found that there are 15468 points with label. I also visualized the point cloud and found it is very different from figure in paper. I think what the code did is keeping one segment of each instance but not keeping only one single point of it. Is this because later you need to propagate the one labled point to the corresponding super voxel(as segment now). So this operation is equivalent to what I described above?

Hi，I have problem is why not just use the features of 3D unet directly? It takes a lot of time to train an extra unet.Have you done any related experiments?

Hi Liu,

In your config file, it has full_scale: [128, 512],

and in the trainMerge(), it has spatial_shape = np.clip((locs.max(0)[0][1:] + 1).numpy(), self.full_scale[0], None)

Could you tell me why do you need to clip the locs with the limitation of 128? Thank you~

Hi @liuzhengzhe!,

I have followed all the steps to successfully install the spconv and import it without any problems, including from spconv import spconv_utils. However, I cannot import spconv inside the model repo (3D-U-Net, merge, relation). The error message is shown below.

     15 import numpy as np
---> 16 from spconv import spconv_utils
     17 from spconv.spconv_utils import (non_max_suppression, non_max_suppression_cpu,
     18                                  points_to_voxel_3d_np, rbbox_iou,

ImportError: cannot import name 'spconv_utils' from 'spconv' (/home/jimmy15923/mnt/project_code/One-Thing-One-Click/3D-U-Net/spconv/__init__.py)

And I can import spconv in the upper repo. Should I remove the original spconv after installation? Thank you

First of all, I really admire your great work, and thanks for sharing your codes!

When I was reading your paper and tracing these codes, I have some questions: In the training phase, you use the confidence of predictions for graph propagation (GP) and replace CRF ( paper section 3.4 ). test_train.py

But in the evaluation phase, you use CRF for GP ( paper ablation study ). test.py

Why does the confidence score appear in the training phase but not in the evaluation phase?

Could you please help clarify the concept behind this difference?

Hi, very cool project!

A quick question: Can you please provide a pointer to the code that calculates the super-voxel partitioning of the input point cloud?

Thank you!

Hello! Thank you for a very nice job! I want to know about the implementation part of the super-voxel pooling, where is the code implementation for this part?

Hi, I am very interested in your work, but there is a problem that bothers me. As mentioned in the article：“In S3DIS, only 36.15 points in each room are annotated on average using “One Thing One Click”, while the original S3DIS has 193797.1 points annotated in each room.” However, as far as I understand， there are nearly 1 million points in each scene on average, so each room should have 200 annotated points(0.02%). Please tell me where is the problem. Thanks!!!

Hi, I ran into a problem when I was deploying spconv. When I run this code cd lib/spconv python setup.py bdist_wheel there is a errorNo CMAKE_CUDA_COMPILER could be found.

Hi, could you please release the data preparation code for scannet data_efficient？

Or, just like otoc, we need to know which 'segments' those labeled data belongs, and then label all the point in these 'segments'?

Hoping your responce sincerely~~

Hi, all. @liuzhengzhe Kindly notice that there is something different in torch.save() between torch1.3 and torch1.6+.

Before doing python prepare_data_otoc.py, should modify line 128, add param_use_new_zipfile_serialization=False Otherwise, torch.load('xxx.pth') will raise an error pytorch_model.bin is a zip archive (did you mean to use torch.jit.load()?)

torch1.3 has no param _use_new_zipfile_serialization=False so while preparing the data, your torch should remain ver1.6 or higher, and after generating .pth files, you should go back to torch1.3 or you will find troubles because PG_OP depends on torch1.3.