Skip to content

Latest commit

 

History

History
165 lines (115 loc) · 8.41 KB

README.md

File metadata and controls

165 lines (115 loc) · 8.41 KB
Raw

2020-02-27

BiSH Segmentation

This is a ROS-compatible implementation of a categorization and semantic segmentation algorithm for point clouds, based on projective analysis. The implementation was developed by Davide Piccinini and Riccardo Monica, at the Robotics and Intelligent Machines Laboratory (RIMLab), Department of Engineering and Architecture, University of Parma, Italy.

The segmentation method was originally described in:

  • Yunhai Wang, Minglun Gong, Tianhua Wang, Daniel Cohen-Or, Hao Zhang, Baoquan Chen, "Projective Analysis for 3D Shape Segmentation", ACM Transactions on Graphics 32, 6, Article 192, November 2013. https://doi.org/10.1145/2508363.2508393

The segmentation algorithm is based on projective analysis, that is:

  • The input (unlabeled) point cloud is projected as if observed by a set of viewpoints. For each viewpoint, a binary image is generated.
  • Each binary image is segmented, and a label is assigned to each pixel.
  • Labels are backprojected on the input point cloud and assigned to each point.

Binary images are segmented by comparison against a dataset of annotated images of objects in the same category. The closest image is found by using the Bi-class Symmetric Hausdorff distance (BiSH). The segmentation is semantically consistent with the annotation in the dataset. Partial code for the BiSH distance was provided by the original authors, downloadable at http://www.yunhaiwang.net/public_html/psa.html.

The algorithm implementation was completed with the full point cloud segmentation pipeline. Moreover, the following features were added:

  • Dataset generation from a set of annotated point clouds in PCD format.
  • Object point cloud categorization.
  • Use of prototype images for faster categorization.
  • ROS integration with action-based interface.

Associated paper:

  • to appear.

Installation

Dependencies:

The repository contains three ROS packages: bish_segmentation_dataset_creation, bish_segmentation and bish_segmentation_msgs. Copy them in your ROS workspace and compile them using catkin build or catkin_make.

Warning: C++ compiler optimization should be enabled for better performance. The package should be compiled in Release or RelWithDebInfo mode.

Dataset generation

The ROS package bish_segmentation_dataset_creation generates a dataset of annotated images from a dataset of annotated point clouds. Point clouds are labeled using the RGB color property: each color corresponds to a different label.

The point clouds must be placed in a sub-folder named labeled_pointclouds in the data folder of the bish_segmentation_dataset_creation package. Point clouds must be split into sub-folders according to their category.

For example, given two categories cups and hammers, the directory structure may be as follows:

- bish_segmentation_dataset_creation
  - data
    - labeled_pointclouds
      - cups
        - cup1.pcd
        - teacup.pcd
        - ... etc.
      - hammers
        - mallet.pcd
        - maul.pcd
        - ... etc.

Then, launch the bish_segmentation_dataset_creation node using the provided launch file:

  roslaunch bish_segmentation_dataset_creation create.launch

This may take some time, depending on the size of your dataset.

After generation, two sub-folders are generated in the data directory: labeled_images and prototypes. labeled_images contains the projected images for segmentation. prototypes contains a subset of labeled_images, selected as prototypes for faster categorization.

The labeled_images and the prototypes folders should be moved to the data folder of the bish_segmentation package, to be used as dataset.

Parameters

  • labeled_pointclouds_path (string): path of the labeled_pointclouds folder.
  • labeled_images_path (string): path of the labeled_images folder.
  • prototypes_path (string): path of the prototypes folder.
  • labeled_segment_number (int): number of horizontal slabs for BiSH segmentation. Increase for more complex objects.
  • point_radius (float): point radius in the input point cloud. Increase if the gap between surfels is visible in the resulting images.
  • normal_k_search_count (int): k-search neighborhood count for normal estimation.
  • image_height (int): generated image height, in pixels.
  • image_width (int): generated image width, in pixels.
  • range_min (float): projection near plane.
  • range_max (float): projection far plane. It should be much larger than object size.
  • theta_x, theta_y, theta_z (float): view sampling steps, in degrees, around the x, y and z axes. Set any of them to 0 to disable rotation around the corresponding axis.
  • prototypes_count (int): number of prototypes to be generated for each class.

Batch classification and segmentation

The ROS package bish_segmentation classifies and segments input unlabeled point clouds. This package requires a dataset in the data folder. The dataset can be generated as explained in the previous section.

In batch mode, the bish_segmentation node classifies and segments multiple PCD files, which are placed in the test_pointclouds sub-folder in the data folder. The directory structure should be as follows:

- bish_segmentation
  - data
    - labeled_images    # dataset
    - prototypes        # dataset
    - test_pointclouds
      - unlabeled_object.pcd
      - unknown_object.pcd
      - ... etc.

Then, launch the bish_segmentation node using the provided launch file:

  roslaunch bish_segmentation classify_and_segment.launch

The node creates several temporary files in data/test_images and data/results. These folders may be safely removed when the node terminates. The segmented files are generated in results_pointclouds.

For each input point cloud, e.g. unknown.pcd, two possible segmented point clouds are generated: unknown-LABELED.pcd and unknown-2_LABELED.pcd. The latter is processed with an additional label optimization step, which should reduce label noise.

Parameters

  • labeled_images_path (string): path of the labeled_images folder.
  • prototypes_path (string): path of the prototypes folder.
  • test_pointclouds_path (string): path of the test_pointclouds folder.
  • test_images_path (string): path of the temporary test_images folder.
  • results_path (string): path of the temporary result folder.
  • results_pointclouds_path (string): path of the results_pointclouds folder.
  • test_segment_number (int): number of horizontal slabs for BiSH segmentation. It should be about twice the labeled_segment_number used during dataset generation.

The following parameters have the same meaning as in bish_segmentation_dataset_creation: point_radius, normal_k_search_count, image_height, image_width, range_min, range_max, theta_x, theta_y, theta_z.

On-demand classification and segmentation

For integration in a robot system, we developed an action-based interface for the bish_segmentation package. The action definition is bish_segmentation_msgs/BishSegment.action.

To use the action-based interface, launch the persistent action_server node:

  roslaunch bish_segmentation action_server.launch

Fields in the action goal:

  • cloud (sensor_msgs/PointCloud2): the point cloud. Points should be of type pcl::PointSurfel. Radius and normal must be provided.
  • view_poses (geometry_msgs/Pose[]): array of poses, a binary image is generated by projection to each pose. If empty, the poses are generated automatically using steps theta_x, theta_y, theta_z.
  • back_face_culling (bool): if true, surfels observed from behind are invisible when generating projections. If the point cloud is incomplete (e.g. obtained from the observation of the object from only one side), this should be set to true.

Fields in the action result:

  • segmented_cloud (sensor_msgs/PointCloud2): segmented cloud, where colors correspond to labels.
  • segmented_class (string): category of the object, corresponding to the sub-folder name during dataset generation.

Note: temporary folders such as data/test_images and data/results are used by the action-based interface as well.

Parameters

All parameters of the bish_segmentation node, and:

  • action_name (string): action name. Default: ~bish_segment.