ML

Open3D-ML is an extension of Open3D for 3D machine learning tasks. It builds on top of the Open3D core library and extends it with machine learning tools for 3D data processing. This repo focuses on applications such as semantic point cloud segmentation and provides pretrained models that can be applied to common tasks as well as pipelines for training.

Open3D-ML works with TensorFlow and PyTorch to integrate easily into existing projects and also provides general functionality independent of ML frameworks such as data visualization.

Installation

Users

Open3D-ML is integrated in the Open3D v0.11+ python distribution and is compatible with the following versions of ML frameworks.

PyTorch 1.7
TensorFlow 2.4
CUDA 10.1, 11.* (On GNU/Linux x86_64, optional)

You can install Open3D with

# make sure you have the latest pip version
pip install --upgrade pip
# install open3d
pip install open3d

To install a compatible version of PyTorch or TensorFlow you can use the respective requirements files:

# To install a compatible version of TensorFlow
pip install -r requirements-tensorflow.txt
# To install a compatible version of PyTorch with CUDA
pip install -r requirements-torch-cuda.txt

To test the installation use

# with PyTorch
$ python -c "import open3d.ml.torch as ml3d"
# or with TensorFlow
$ python -c "import open3d.ml.tf as ml3d"

If you need to use different versions of the ML frameworks or CUDA we recommend to build Open3D from source.

Getting started

Reading a dataset

The dataset namespace contains classes for reading common datasets. Here we read the SemanticKITTI dataset and visualize it.

import open3d.ml.torch as ml3d  # or open3d.ml.tf as ml3d

# construct a dataset by specifying dataset_path
dataset = ml3d.datasets.SemanticKITTI(dataset_path='/path/to/SemanticKITTI/')

# get the 'all' split that combines training, validation and test set
all_split = dataset.get_split('all')

# print the attributes of the first datum
print(all_split.get_attr(0))

# print the shape of the first point cloud
print(all_split.get_data(0)['point'].shape)

# show the first 100 frames using the visualizer
vis = ml3d.vis.Visualizer()
vis.visualize_dataset(dataset, 'all', indices=range(100))

Loading a config file

Configs of models, datasets, and pipelines are stored in ml3d/configs. Users can also construct their own yaml files to keep record of their customized configurations. Here is an example of reading a config file and constructing modules from it.

import open3d.ml as _ml3d
import open3d.ml.torch as ml3d # or open3d.ml.tf as ml3d

framework = "torch" # or tf
cfg_file = "ml3d/configs/randlanet_semantickitti.yml"
cfg = _ml3d.utils.Config.load_from_file(cfg_file)

# fetch the classes by the name
Pipeline = _ml3d.utils.get_module("pipeline", cfg.pipeline.name, framework)
Model = _ml3d.utils.get_module("model", cfg.model.name, framework)
Dataset = _ml3d.utils.get_module("dataset", cfg.dataset.name)

# use the arguments in the config file to construct the instances
cfg.dataset['dataset_path'] = "/path/to/your/dataset"
dataset = Dataset(cfg.dataset.pop('dataset_path', None), **cfg.dataset)
model = Model(**cfg.model)
pipeline = Pipeline(model, dataset, **cfg.pipeline)

Semantic Segmentation

Running a pretrained model for semantic segmentation

Building on the previous example we can instantiate a pipeline with a pretrained model for semantic segmentation and run it on a point cloud of our dataset. See the model zoo for obtaining the weights of the pretrained model.

import os
import open3d.ml as _ml3d
import open3d.ml.torch as ml3d

cfg_file = "ml3d/configs/randlanet_semantickitti.yml"
cfg = _ml3d.utils.Config.load_from_file(cfg_file)

model = ml3d.models.RandLANet(**cfg.model)
cfg.dataset['dataset_path'] = "/path/to/your/dataset"
dataset = ml3d.datasets.SemanticKITTI(cfg.dataset.pop('dataset_path', None), **cfg.dataset)
pipeline = ml3d.pipelines.SemanticSegmentation(model, dataset=dataset, device="gpu", **cfg.pipeline)

# download the weights.
ckpt_folder = "./logs/"
os.makedirs(ckpt_folder, exist_ok=True)
ckpt_path = ckpt_folder + "randlanet_semantickitti_202009090354utc.pth"
randlanet_url = "https://storage.googleapis.com/open3d-releases/model-zoo/randlanet_semantickitti_202009090354utc.pth"
if not os.path.exists(ckpt_path):
    cmd = "wget {} -O {}".format(randlanet_url, ckpt_path)
    os.system(cmd)

# load the parameters.
pipeline.load_ckpt(ckpt_path=ckpt_path)

test_split = dataset.get_split("test")
data = test_split.get_data(0)

# run inference on a single example.
# returns dict with 'predict_labels' and 'predict_scores'.
result = pipeline.run_inference(data)

# evaluate performance on the test set; this will write logs to './logs'.
pipeline.run_test()

Users can also use predefined scripts to load pretrained weights and run testing.

Training a model for semantic segmentation

Similar as for inference, pipelines provide an interface for training a model on a dataset.

# use a cache for storing the results of the preprocessing (default path is './logs/cache')
dataset = ml3d.datasets.SemanticKITTI(dataset_path='/path/to/SemanticKITTI/', use_cache=True)

# create the model with random initialization.
model = RandLANet()

pipeline = SemanticSegmentation(model=model, dataset=dataset, max_epoch=100)

# prints training progress in the console.
pipeline.run_train()

For more examples see examples/ and the scripts/ directories.

3D Object Detection

Running a pretrained model for 3D object detection

The 3D object detection model is similar to a semantic segmentation model. We can instantiate a pipeline with a pretrained model for Object Detection and run it on a point cloud of our dataset. See the model zoo for obtaining the weights of the pretrained model.

import os
import open3d.ml as _ml3d
import open3d.ml.torch as ml3d

cfg_file = "ml3d/configs/pointpillars_kitti.yml"
cfg = _ml3d.utils.Config.load_from_file(cfg_file)

model = ml3d.models.PointPillars(**cfg.model)
cfg.dataset['dataset_path'] = "/path/to/your/dataset"
dataset = ml3d.datasets.KITTI(cfg.dataset.pop('dataset_path', None), **cfg.dataset)
pipeline = ml3d.pipelines.ObjectDetection(model, dataset=dataset, device="gpu", **cfg.pipeline)

# download the weights.
ckpt_folder = "./logs/"
os.makedirs(ckpt_folder, exist_ok=True)
ckpt_path = ckpt_folder + "pointpillars_kitti_202012171738utc.pth"
pointpillar_url = "https://storage.googleapis.com/open3d-releases/model-zoo/pointpillars_kitti_202012171738utc.pth"
if not os.path.exists(ckpt_path):
    cmd = "wget {} -O {}".format(pointpillar_url, ckpt_path)
    os.system(cmd)

# load the parameters.
pipeline.load_ckpt(ckpt_path=ckpt_path)

test_split = dataset.get_split("test")
data = test_split.get_data(0)

# run inference on a single example.
# returns dict with 'predict_labels' and 'predict_scores'.
result = pipeline.run_inference(data)

# evaluate performance on the test set; this will write logs to './logs'.
pipeline.run_test()

Users can also use predefined scripts to load pretrained weights and run testing.

Training a model for 3D object detection

Similar as for inference, pipelines provide an interface for training a model on a dataset.

# use a cache for storing the results of the preprocessing (default path is './logs/cache')
dataset = ml3d.datasets.KITTI(dataset_path='/path/to/KITTI/', use_cache=True)

# create the model with random initialization.
model = PointPillars()

pipeline = ObjectDetection(model=model, dataset=dataset, max_epoch=100)

# prints training progress in the console.
pipeline.run_train()

Below is an example of visualization using KITTI. The example shows the use of bounding boxes for the KITTI dataset.

For more examples see examples/ and the scripts/ directories.

Using predefined scripts

scripts/run_pipeline.py provides an easy interface for training and evaluating a model on a dataset. It saves the trouble of defining specific model and passing exact configuration.

python scripts/run_pipeline.py {tf/torch} -c <path-to-config> --pipeline {SemanticSegmentation/ObjectDetection} --<extra args>

You can use script for both semantic segmentation and object detection. You must specify either SemanticSegmentation or ObjectDetection in the pipeline parameter. Note that extra args will be prioritized over the same parameter present in the configuration file. So instead of changing param in config file, you may pass the same as a command line argument while launching the script.

For eg.

# Launch training for RandLANet on SemanticKITTI with torch.
python scripts/run_pipeline.py torch -c ml3d/configs/randlanet_semantickitti.yml --dataset.dataset_path <path-to-dataset> --pipeline SemanticSegmentation --dataset.use_cache True

# Launch testing for PointPillars on KITTI with torch.
python scripts/run_pipeline.py torch -c ml3d/configs/pointpillars_kitti.yml --split test --dataset.dataset_path <path-to-dataset> --pipeline ObjectDetection --dataset.use_cache True

For further help, run python scripts/run_pipeline.py --help.

Repository structure

The core part of Open3D-ML lives in the ml3d subfolder, which is integrated into Open3D in the ml namespace. In addition to the core part, the directories examples and scripts provide supporting scripts for getting started with setting up a training pipeline or running a network on a dataset.

├─ docs                   # Markdown and rst files for documentation
├─ examples               # Place for example scripts and notebooks
├─ ml3d                   # Package root dir that is integrated in open3d
     ├─ configs           # Model configuration files
     ├─ datasets          # Generic dataset code; will be integratede as open3d.ml.{tf,torch}.datasets
     ├─ metrics           # Metrics available for evaluating ML models
     ├─ utils             # Framework independent utilities; available as open3d.ml.{tf,torch}.utils
     ├─ vis               # ML specific visualization functions
     ├─ tf                # Directory for TensorFlow specific code. same structure as ml3d/torch.
     │                    # This will be available as open3d.ml.tf
     ├─ torch             # Directory for PyTorch specific code; available as open3d.ml.torch
          ├─ dataloaders  # Framework specific dataset code, e.g. wrappers that can make use of the
          │               # generic dataset code.
          ├─ models       # Code for models
          ├─ modules      # Smaller modules, e.g., metrics and losses
          ├─ pipelines    # Pipelines for tasks like semantic segmentation
          ├─ utils        # Utilities for <>
├─ scripts                # Demo scripts for training and dataset download scripts

Tasks and Algorithms

Semantic Segmentation

For the task of semantic segmentation, we measure the performance of different methods using the mean intersection-over-union (mIoU) over all classes. The table shows the available models and datasets for the segmentation task and the respective scores. Each score links to the respective weight file.

Model / Dataset	SemanticKITTI	Toronto 3D	S3DIS	Semantic3D	Paris-Lille3D	ScanNet
RandLA-Net (tf)	53.7	69.0	67.0	76.0	70.0	-
RandLA-Net (torch)	52.8	71.2	67.0	76.0	70.0	-
KPConv (tf)	58.7	65.6	65.0	-	76.7	-
KPConv (torch)	58.0	65.6	60.0	-	76.7	-
SparseConvUnet (torch)	-	-	-	-	-	68
SparseConvUnet (tf)	-	-	-	-	-	68.2

Object Detection

For the task of object detection, we measure the performance of different methods using the mean average precision (mAP) for bird's eye view (BEV) and 3D. The table shows the available models and datasets for the object detection task and the respective scores. Each score links to the respective weight file. For the evaluation, the models were evaluated using the validation subset, according to KITTI's validation criteria. The models were trained for three classes (car, pedestrian and cyclist). The calculated values are the mean value over the mAP of all classes for all difficulty levels.

Model / Dataset	KITTI [BEV / 3D] @ 0.70
PointPillars (tf)	61.6 / 55.2
PointPillars (torch)	61.2 / 52.8
PointRCNN (tf)	78.2 / 65.9
PointRCNN (torch)	78.2 / 65.9

Training PointRCNN

To use ground truth sampling data augmentation for training, we can generate the ground truth database as follows:

python scripts/collect_bboxes.py --dataset_path <path_to_data_root>

This will generate a database consisting of objects from the train split. It is recommended to use this augmentation for dataset like KITTI where objects are sparse.

The two stages of PointRCNN are trained separately. To train the proposal generation stage of PointRCNN with PyTorch, run the following command:

# Train RPN for 100 epochs.
python scripts/run_pipeline.py torch -c ml3d/configs/pointrcnn_kitti.yml --dataset.dataset_path <path-to-dataset> --mode RPN --epochs 100

After getting a well trained RPN network, we can train RCNN network with frozen RPN weights.

# Train RCNN for 70 epochs.
python scripts/run_pipeline.py torch -c ml3d/configs/pointrcnn_kitti.yml --dataset.dataset_path <path-to-dataset> --mode RCNN --model.ckpt_path <path_to_checkpoint> --epochs 100

Model Zoo

For a full list of all weight files see model_weights.txt and the MD5 checksum file model_weights.md5.

Datasets

The following is a list of datasets for which we provide dataset reader classes.

SemanticKITTI (project page)
Toronto 3D (github)
Semantic 3D (project-page)
S3DIS (project-page)
Paris-Lille 3D (project-page)
Argoverse (project-page)
KITTI (project-page)
Lyft (project-page)
nuScenes (project-page)
Waymo (project-page)
ScanNet(project-page)

For downloading these datasets visit the respective webpages and have a look at the scripts in scripts/download_datasets.

How-tos

Visualize network predictions
Visualize custom data
Adding a new model
Adding a new dataset

Contribute

There are many ways to contribute to this project. You can:

Implement a new model
Add code for reading a new dataset
Share parameters and weights for an existing model
Report problems and bugs

Please, make your pull requests to the dev branch. Open3D is a community effort. We welcome and celebrate contributions from the community!

If you want to share weights for a model you trained please attach or link the weights file in the pull request. For bugs and problems, open an issue. Please also check out our communication channels to get in contact with the community.

Communication channels

Forum: discussion on the usage of Open3D.
Discord Chat: online chats, discussions, and collaboration with other users and developers.

Citation

Please cite our work (pdf) if you use Open3D.

@article{Zhou2018,
    author    = {Qian-Yi Zhou and Jaesik Park and Vladlen Koltun},
    title     = {{Open3D}: {A} Modern Library for {3D} Data Processing},
    journal   = {arXiv:1801.09847},
    year      = {2018},
}

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Installation

Users

Getting started

Reading a dataset

Loading a config file

Semantic Segmentation

Running a pretrained model for semantic segmentation

Training a model for semantic segmentation

3D Object Detection

Running a pretrained model for 3D object detection

Training a model for 3D object detection

Using predefined scripts

Repository structure

Tasks and Algorithms

Semantic Segmentation

Object Detection

Training PointRCNN

Model Zoo

Datasets

How-tos

Contribute

Communication channels

Citation

Files

README.md

Latest commit

History

README.md

File metadata and controls

Installation

Users

Getting started

Reading a dataset

Loading a config file

Semantic Segmentation

Running a pretrained model for semantic segmentation

Training a model for semantic segmentation

3D Object Detection

Running a pretrained model for 3D object detection

Training a model for 3D object detection

Using predefined scripts

Repository structure

Tasks and Algorithms

Semantic Segmentation

Object Detection

Training PointRCNN

Model Zoo

Datasets

How-tos

Contribute

Communication channels

Citation