A fast, memory efficient free and open source point cloud classifier. It generates an AI model from a set of input point clouds that have been labeled and can subsequently use that model to classify new datasets.
On the default parameters it can classify 15 million points in less than 2 minutes on a 4-core Intel i5, which is faster than any other freely available software we've tried.
It generalizes well to point clouds of varying density and includes local smoothing regularization methods.
It supports all point cloud formats supported by PDAL. When built without PDAL, it supports a subset of the PLY format only, which is optimized for speed.
If you're on Windows, you have two choices:
- Purchase the program for Windows and support the project ❤️
- Build the program from sources (see instructions below)
If you're on macOS/Linux you currently need to build from sources (see instructions below).
Dependencies:
- Intel TBB
- Eigen
- PDAL (optional for LAS/LAZ support)
mkdir build
cd build
cmake .. [-DWITH_GBT=ON]
make -j$(nproc)
You will need Visual Studio, CMake and VCPKG.
Install required packages with VCPKG:
vcpkg install eigen3 tbb pdal
Replace <VCPKG_PATH> with the path to your VCPKG installation in the following commands:
mkdir build
cd build
cmake [-DWITH_GBT=ON] -DCMAKE_TOOLCHAIN_FILE=<VCPKG_PATH>/scripts/buildsystems/vcpkg.cmake ..
cmake --build . --config Release --target ALL_BUILD -- /maxcpucount:14
./pctrain ./ground_truth.ply
./pcclassify ./dataset.ply ./classified.ply [model.bin]
We provide access to a pre-trained model if you don't have access to labeled data. Please note the model was generated using a limited number of samples and it might not work well with all datasets.
Training classes are assumed to follow the ASPRS 1.4 Classification and to be stored in either a label, class or classification property.
You can re-map classification codes by creating a <FILE>.json in the same directory as <FILE>.ply:
{
"source": "https://url-to-your-data",
"classification": {
"0": "ground",
"1": "building",
"2": "low_vegetation",
"3": "medium_vegetation",
"4": "high_vegetation",
"17": "ground"
}
}
You can also use the --classes flag to limit training to a subset of the ASPRS classes:
./pctrain ./ground_truth.laz --classes 2,5,6
| Class | Number |
|---|---|
| unclassified | 1 |
| ground | 2 |
| low_vegetation | 3 |
| medium_vegetation | 4 |
| high_vegetation | 5 |
| building | 6 |
| low_point | 7 |
| water | 9 |
| rail | 10 |
| road_surface | 11 |
| wire_guard | 13 |
| wire_conductor | 14 |
| transmission_tower | 15 |
| wire_structure_connector | 16 |
| bridge_deck | 17 |
| high_noise | 18 |
| overhead_structure | 19 |
| ignored_ground | 20 |
| snow | 21 |
| temporal_exclusion | 22 |
| human_made_object | 64 |
You can check a model accuracy by using the --eval argument:
./pctrain ./ground_truth.ply --eval test.ply
You can use PDAL to conveniently split a dataset into two (one for training, one for evaluation):
pdal split [--capacity numpoints] input.ply input_split.ply
You can output the results of classification as a colored point cloud by using the --color option:
./pcclassify ./dataset.ply ./classified.ply --color
pctrain can generate AI models using either random forests (default) or gradient boosted trees:
./pctrain -c gbt [...]
See ./pctrain --help.
You can build a Docker image with the following command:
docker build -t uav4geo/openpointclass:latest .Run the image with the following command:
docker run -it --rm -v /dataset-path:/data uav4geo/openpointclass:latest bashWhere /dataset-path is the path to the directory containing the dataset files and the model.bin file.
You will be presented with a bash prompt inside the container. You can then run the pctrain and pcclassify as described above.
Otherwise, you can use the commands directly with the following syntax:
docker run -it --rm -v /dataset-path:/data uav4geo/openpointclass:latest pctrain /data/ground_truth.ply
docker run -it --rm -v /dataset-path:/data uav4geo/openpointclass:latest pcclassify /data/dataset.ply /data/classified.ply /data/model.bin- We only support a subset of the PLY format (for performance reasons) and certain less common variations of the format might give trouble. Most importantly, the X/Y/Z coordinates must be
floatvalues when using binary PLY, notdoubleorfloat64. We recommend to use LAS/LAZ if higher precision coordinates are needed.
The software is released under the terms of the AGPLv3
Made with ❤️ by UAV4GEO
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent