Diffuser

This repository contains the implementation of the multi-view semantic scene labeling pipeline described in:

Ruben Mascaro, Lucas Teixeira, and Margarita Chli. Diffuser: Multi-View 2-D to 3-D Label Diffusion for Semantic Scene Segmentation. IEEE International Conference on Robotics and Automation (ICRA), 2021.

Taking as input a 3D point cloud and a set of localized images processed by a 2D semantic segmentation network, Diffuser uses an efficient graphical model that leverages geometry to propagate class labels from the 2D image space to the 3D map. For more specific details, please check our [Paper] and [Video].

Environment Setup

This package requires Python 3.7+ and depends on the following libraries:

• CuPy
• NumPy
• Open3D
• Pillow
• SciPy

We recommend setting up a conda environment, as it simplifies the installation of the CUDA runtime libraries required by CuPy. A working conda environment can be created using the provided environment-gpu.yml file:

conda env create -f environment-gpu.yml
conda activate dfsr

NOTE: We also provide a CPU-only implementation that does not require CuPy and can be run on machines without a NVIDIA GPU. To set up a conda environment without support for GPU-accelerated computing, please use the environment-cpu.yml file:

conda env create -f environment-cpu.yml
conda activate dfsr

Example Usage on ScanNet

Dataset Preparation

Request access to the ScanNet data following the instructions in the ScanNet repository. Upon acceptance of your request, you will be sent a link to a download script called download-scannet.py. Please copy this script into the tools/datasets/ directory.

Then, clone the ScanNet repository and build the ScanNet C++ Toolkit (this is required to extract the RGB-D frames and camera intrinsics and extrinsics from the compressed ScanNet .sens data):

git clone git@github.com:ScanNet/ScanNet.git
cd ScanNet/SensReader/c++/
make

Once the previous steps are completed, you can use the provided download_scannet_scene.sh script to automatically download and preprocess any ScanNet scene. Example:

./tools/datasets/download_scannet_scene.sh scene0694_00 data/scannet ../ScanNet

Running Diffuser

First, process the input frames with your semantic segmentation network of choice (e.g. MSeg, EMSANet, etc.).

Then, set up your own config file in the config directory, making sure that the input data paths point to the desired directories. Note that we provide an example config file (scannet.ini) that can be easily modified.

Finally, execute the main.py script as:

python main.py configs/scannet.ini

Adding New Datasets

To test Diffuser on a custom dataset, please follow these steps:

1. Create a new file diffuser/datasets/mydataset.py.

from .base_dataset import BaseLabelTaxonomy, BaseScene
from .builder import (register_dataset, register_label_mapping,
                      register_label_taxonomy)


# Adding a custom label taxonomy:
# Here you just need to define the class names and colors used for 3D semantic 
# labeling. The codebase assumes that the label of each class corresponds to 
# its position in the CLASSES array (e.g. 'class0' has label 0, class1' has 
# label 1, and so on).
@register_label_taxonomy('mylabeltaxonomy')
class MyLabelTaxonomy(BaseLabelTaxonomy):
  
  CLASSES = ('class0', 'class1', ...)
  PALETTE = ((r0, g0, b0), (r1, g1, b1), ...)

  def __init__(self):
        super().__init__()


# Adding a custom label mapping:
# Label mappings are needed if your semantic segmentation network of choice 
# uses a label taxonomy that does not exactly match the desired taxonomy for 3D 
# semantic labeling. The function should return a mapping from the input image 
# labels to the desired labels in 3D. In the example below, labels x and y in 
# image space would be mapped to label 0, while label z in image space would be
# mapped to label 1.  
@register_label_mapping('mylabelmapping')
def my_label_mapping():
  return [[x, y], [z], ...]


# Adding a custom dataset:
# To add a new dataset, you must provide the initialization function as well as
# the implementation of the abstract methods for data loading in the BaseScene 
# class. See `diffuser/datasets/scannet.py` for an example.
@register_dataset('scannet')
class MyDatasetScene(BaseScene):
  
  def __init__(self, arg1, arg2, ...):
        ...

2. Import the module in diffuser/datasets/__init__.py.

from .mydataset import MyDatasetScene, MyLabelTaxonomy

3. Use the dataset by creating a new config file configs/mydataset.ini.

[dataset]
dataset_name = mydataset
label_taxonomy = mylabeltaxonomy
img_labels_mapping = mylabelmapping
...

License

This project is released under the BSD 3-Clause License.

Citation

If you use this code in your academic work, please consider citing:

@inproceedings{mascaro2021diffuser,
  title={Diffuser: Multi-View 2-D to 3-D Label Diffusion for Semantic Scene Segmentation},
  author={Mascaro, Ruben and Teixeira, Lucas and Chli, Margarita},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2021}
}