Cloudini (pronounced with Italian accent) is a pointcloud compression library.

Its main focus is speed, but it still achieves very good compression ratios.

Its main use cases are:

• To improve the storage of datasets containing pointcloud data (being a notable example rosbags).

• Decrease the bandwidth used when streaming pointclouds over a network.

It works seamlessly with PCL and ROS, but the main library can be compiled and used independently, if needed.

What to expect

The compression ratio is hard to predict because it depends on the way the original data is encoded.

For example, ROS pointcloud messages are extremely inefficient, because they include some "padding" in the message that, in extreme cases, may reach up to 50%.

(Yes, you heard correctly, almost 50% of that 10 Gb rosbag is useless padding).

But, in general, you may expect considerably better compression and faster encoding/decoding than ZSTD or LZ4 alone.

These are two random examples using real-world data from LiDARs.

• Channels: XYZ, Intensity, no padding

  [LZ4 only]      ratio: 0.77 time (usec): 2165
  [ZSTD only]     ratio: 0.68 time (usec): 2967
  [Cloudini-LZ4]  ratio: 0.56 time (usec): 1254
  [Cloudini-ZSTD] ratio: 0.51 time (usec): 1576

• Channels: XYZ, intensity, ring (int16), timestamp (double), with padding

  [LZ4 only]      ratio: 0.31 time (usec): 2866
  [ZSTD only]     ratio: 0.24 time (usec): 3423
  [Cloudini-LZ4]  ratio: 0.16 time (usec): 2210
  [Cloudini-ZSTD] ratio: 0.14 time (usec): 2758

If you are a ROS user, you can test the compression ratio and speed yourself, running the application rosbag_benchmark on any rosbag containing a sensor_msgs::msg::PointCloud2 topic.

How to test it yourself

There is a pre-compiled Linux AppImage that can be downloaded in the release page

Alternatively, you can test the obtainable compression ratio in your browser here: https://cloudini.netlify.app/

NOTE: your data will not be uploaded to the cloud. The applications runs 100% inside your browser.

How it works

The algorithm contains two steps:

1. Encoding the pointcloud, channel by channel.
2. Compression using either LZ4 or ZSTD.

The encoding is lossy for floating point channels (typically the X, Y, Z channels) and lossless for RGBA and integer channels.

Now, I know that when you read the word "lossy" you may think about grainy JPEGS images. Don't.

The encoder applies a quantization using a resolution provided by the user.

Typical LiDARs have an accuracy/noise in the order of +/- 2 cm. Therefore, using a resolution of 1 mm (+/- 0.5 mm max quantization error) is usually a very conservative option.

But, if you are really paranoid, and decide to use a resolution of 100 microns, you still achieve excellent compression ratios!

It should also be noted that this two-step compression strategy has a negative overhead, i.e. it is actually faster than using LZ4 or ZSTD alone.

Compile instructions

Some dependencies are downloaded automatically using CPM. To avoid downloading them again when your rebuild your project, I suggest setting CPM_SOURCE_CACHE as described here.

To build the main library (cloudini_lib)

cmake -B build -S cloudini_lib -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel

To compile it with ROS, just pull this repo into your ws/src folder and execute colcon build as usual.

ROS specific utilities

For more information, see the cloudini_ros/README.md

• point_cloud_transport plugins: see point_cloud_transport plugins for reference about how they are used.

• cloudini_topic_converter: a node that subscribes to a compressed point_cloud_interfaces/CompressedPointCloud2 and publishes a sensor_msgs/PointCloud2.

• cloudini_rosbag_converter: a command line tool that, given a rosbag (limited to MCAP format), converts all sensor_msgs/PointCloud2 topics into compressed point_cloud_interfaces/CompressedPointCloud2 of vice-versa.

Frequently Asked Questions

How does it perform, compared to Draco?

Google Draco has two main encoding methods: SEQUENTIAL and KD_TREE.

The latter could achieve excellent compression ratios, but it is very sloooow and it doesn't preserve the original order of the points in the point cloud.

Compared with the Draco sequential mode, Cloudini achieves approximately the same compression, but is considerably faster in my (currently limited) benchmark.

Does the decoder need to know if LZ4 or ZSTD was used?

No, that information is stored in the header of the compressed data, and the decoder will automatically select the right library.