By Peiyun Hu, David Held, and Deva Ramanan at Carnegie Mellon University.

For segmenting LiDAR point clouds, if we score a segmentation by the worst objectness score among its individual segments, there is an algorithm that efficiently finds the optimal worst-case segmentation among an exponentially large number of candidate segmentations. The proposed algorithm takes a pre-processed LIDAR point cloud (top - with background removed) and produces a class-agnostic instance-level segmentation over all foreground points (bottom). We use a different color for each segment and plot an extruded polygon to show the spatial extent.

You can read our paper (open-access) here: https://ieeexplore.ieee.org/abstract/document/8954778.

In this repo, we provide our implementation of this work.

If you find our work useful in your research, please consider citing:

@article{hu2020learning,
  title={Learning to Optimally Segment Point Clouds},
  author={Hu, Peiyun and Held, David and Ramanan, Deva},
  journal={IEEE Robotics and Automation Letters},
  year={2020},
  publisher={IEEE}
}

Currently, code release is a work in progress. Below are what I plan to work on next:

• Update README to describe
  • How to train the objectness model (PointNets) (pointnet2/)
  • How to run segmentation (segment_with_pointnet.py)
  • How to evaluate under-segmentation and over-segmentation (evaluate_under_over*.py)
  • How to evaluate instance-segmentation (evaluate_instance_all.py)
  • How to evaluate existing detectors (evaluate_instance_all.py)
• Merge evaluate_under_over.py and evaluate_under_over_ovlp_part_ignored.py
  • When evaluating under-segmentation and over-segmentation, we either
    • Ignore objects with overlapping bounding boxes
    • Or ignore points that fall into the overlapping regions
  • Right now, they are highly redundant. I plan to merge them together.
• Release all pre-trained models.
• Test training code.