dataloader for TUM RGB-D SLAM Datasets about kiss-icp

[nachovizzo]

Hello, thanks for the interest in our project.

I just made a draft of the dataloader, but there is still some details to fix. Could you please give it a try and provide some feedback to see if it's working?

How to try it?

git clone  https://github.com/PRBonn/kiss-icp.git
cd kiss-icp
git checkout nacho/add_tum_dataloader

After that, you can ran pip install --verbose . to install this branch

How to run?

kiss_icp_pipeline --dataloader tum $DATASETS/tum/rgbd_dataset_freiburg1_xyz --visualize --max_range 4.0

Expected output

The evaluation is still wrong but we are working on finishing the draft of the dataloader.

[lsorgi]

hello Ignazio, many thanks for your prompt answer.
and thanks for making all this opensource. great work :-)

[lsorgi]

feedback about installation on Windows laptop:

the command
pip install --verbose .
fails with the following error

ERROR: Failed building wheel for kiss-icp
  Failed to build kiss-icp
  ERROR: Could not build wheels for kiss-icp, which is required to install pyproject.toml-based projects

the command
pip install -e ".[all]"
completes successfully

[lsorgi]

I tried to run the command given above and it raises and error
ModuleNotFoundError: No module named 'kiss_icp.pybind'

[nachovizzo]

That means that you couldn't build the project, can you paste the entire log? Additionally you can also patch the pyproject.toml since there is a bug in setuptools introduced some hours ago

commit 4e9cab597d22542b0a877ecebde0a234be50443b
Author: Ignacio Vizzo <[email protected]>
Date:   Mon Jan 16 16:36:25 2023 +0100

    Fix editable mode installation

diff --git a/pyproject.toml b/pyproject.toml
index da3a4bb..cda1725 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -3,7 +3,7 @@ requires = [
     "cmake",
     "ninja",
     "scikit-build",
-    "setuptools",
+    "setuptools<66.0.0", # https://github.com/pypa/setuptools/issues/3772
 ]
 build-backend = "setuptools.build_meta"
 

[lsorgi]

I tried to patch the pyproject.toml file and I get the same error. (trying now to build on Ubuntu 18.04)
Here is the full log from Windows.

setup_log.txt

[nachovizzo]

Then try with pip install -U cmake ninja

[lsorgi]

I tried installing ninja but I get errors on both Windows and Ubuntu (see attached files)

Windows_setup_log.txt
Ubuntu1804_setup_log.txt

[nachovizzo]

pip version? Are you using conda or any other animal alike? The problem is that the build system is not finding ninja on the path.

[lsorgi]

yes using similar animal, miniconda

[nachovizzo]

Ah ok, then I can't help with that, since I'm allergic to those animals :P they never really worked for me :D.

kiss-icp does not have many dependencies, so there is actually no big gain on isolating the whole thing ;)

[lsorgi]

I'll do a few more attempts within the conda env.

• I think I managed to install correctly ninja (https://anaconda.org/conda-forge/ninja)
• the build error is different now, maybe you see something?
  Ubuntu1804_setup_log.txt

[nachovizzo]

I see the same error:

'/tmp/pip-build-env-d5sl4e3f/overlay/bin/ninja' '--version'

    failed with:

     No such file or directory

Now you know why I don't like conda :p. Too complicated to manage just some python dependcies 😇

Have you tried outside conda?

[lsorgi]

issue fixed!
g++-9 is needed, nothing to do with anaconda. you may want to include this in the readme.
what I did is

• install g++-9
• update symbolic link g++ -> g++-9 executable
  and the build was successful

[nachovizzo]

Great! It's actually already on the Readme. But I don't want to enforce any particular version, since it depends on your OS

[lsorgi]

I am trying to process the TUM dataset in a simple loop but the final map does not match the one showed above.
Below is pseudo code & the KISS configuration. Anything wrong in the way I am using the library?

versor_map = ...  # this is HW x 3 matrix containing the 3d unit norm vector corresponding to each pixel 

for filename in depth_file_list:
          logging.info(f'Processing data file ({filename})')

          # extract timestamp from filename
         timestamp, _ = os.path.splitext(os.path.basename(filename))
         timestamp = float(timestamp)
    
         # load depth map
         depth_map = cv2.imread(filename, cv2.IMREAD_ANYDEPTH)

        # convert to 3D data        
        depth_map = depth_map.flatten() * 1e-2  # convert to [cm]
        flags = depth_map > 0  # exclude invalid depth (depth == 0)
        pts_3d = versor_map[flags, :] * np.tile(depth_map[flags, np.newaxis], reps=(1, 3))

        # run SLAM
        odometry.register_frame(frame=pts_3d, timestamps=timestamp)




      KISS-ICP Configuration: 
      {
          'adaptive_threshold': 
          {
               'fixed_threshold': 0.3,
               'initial_threshold': 1.0,
               'min_motion_th': 0.01
          },
          'data': 
           {
                 'correct_scan': True,
                 'lidar_frequency': 10.0,
                 'max_range': 100.0,
                 'min_range': 0.0,
                'preprocess': True
          },
         'mapping': 
         {
              'max_points_per_voxel': 20, 
              'voxel_size': 1.0
         },
         'out_dir': 'results'
 }

[nachovizzo]

@lsorgi All the development for the TUM dataset is happening at #45 . I suggest you stay tunned for when we merge this into the main branch so you can give it a try

[nachovizzo]

@lsorgi we just merged #45 please let us know if everything is working on your side!