Pointcloud Format

[Nik-V9]

Hi, Thanks for sharing this dataset!

I'm trying to load the KITTI format pointclouds. When I use the commands shared in the README, I'm getting a numpy array with many dtypes.

I tried doing the following, however the visualization of the pointcloud (pts) looks incorrect:

pts_data = np.fromfile(bin_path, '<f4')
pts = np.reshape(pts_data, (-1, 4))
pts = pts[:, :3]

I was wondering what's the correct way to load in the LiDAR data? Also, do you guys have a PyTorch dataloader or an example?

[EmanueleGiacomini]

Hi, thank you for your interest :smile: Compared to original KITTI clouds that express every point as a 4-dimensional float32 vector <x, y, z, intensity>, We store for every point the complete set of channels exposed by our sensors <x, y, z, intensity, t, reflectivity, ring, ambient, range> with their respective datatypes <f4, f4, f4, f4, u4, u2, u1, u2, u4>.

During the conversion, we expose the appropriate datatype that users can use to parse the clouds. I will provide here an extended code snipper to get clouds in the typical XYZI format:

import numpy as np
import pickle

with open("campus_train0_00_kitti/ouster_points/data/.dtype.pkl", "rb") as f:
    cdtype = pickle.load(f)

cloud_np = np.fromfile("/campus_train0_00_kitti/ouster_points/data/0000000000.bin", dtype=cdtype)
pts = np.stack([
  cloud_np["x"],
  cloud_np["y"],
  cloud_np["z"],
  cloud_np["intensity"]], axis=1)

In this way, pts will contain data properly aligned as typical KITTI clouds

$> pts.shape
> (65536, 4)

You can also visualize the full set of channels by printing cdtype

$> print(cdtype)
> [('x', '<f4'), ('y', '<f4'), ('z', '<f4'), ('intensity', '<f4'), ('t', '<u4'), ('reflectivity', '<u2'), ('ring', 'u1'), ('ambient', '<u2'), ('range', '<u4')]

Concerning the PyTorch dataloader, if you provide us with more details, we can either extend our package or provide the appropriate snippets :smile:

[Nik-V9]

Thanks for sharing this!

Re. PyTorch dataloader: I'm implementing a dataloader which can load in the camera images and the associated poses and depth maps. I'll open more issues here if I'm unable to figure things out.

[EmanueleGiacomini]

For anyone looking forward to this issue, we included support from version 0.0.9 to convert our PointClouds to the same format as KITTI clouds. During the conversion, use the flag --reduce-pcloud to preserve only the <x, y, z, intensity> components stored as float32 elements. This allows previous KITTI readers to parse our clouds.