Live Reconstruction of Large-Scale Dynamic Outdoor Worlds

ABOUT

This code implements the state-of-the-art system for dense 3D reconstruction of outdoor environments described in the CVPR 2019 workshop paper Live Reconstruction of Large-Scale Dynamic Outdoor Worlds by Ondrej Miksik and Vibhav Vineet.

@inproceedings{miksik2019cvprw,
  author = {Ondrej Miksik and Vibhav Vineet},
  title = {{L}ive {R}econstruction of {L}arge-{S}cale {D}ynamic {O}utdoor {W}orlds},
  booktitle = {{CVPR} {W}orkshop on {D}ynamic {S}cene {R}econstruction},
  year = {2019}
}

Contacts

For any questions about the code or the paper, feel free to contact us. More info can be found on the project page: http://www.miksik.co.uk/projects/dfusion/dfusion.html

HOW TO RUN

Prerequisites

This project is written in Python 3 and requires Open3D and standard packages like Matplotlib Pandas or NumPy.

Data / How can I process my sequences?

We mostly use virtual kitti format, see examples in data folder and details in the paper.

Examples

1. dynamic reconstruction

  python main.py --dataset kitti_0003 --d 0 --r --c --f

Optional parameters:

• --start 10 - reconstruction will start from 10th frame

• --limit_frames 20 - will reconstruct only 20 frames

• --no-c - object instances will not be colour coded

• --o - enables on-the-fly output (storing of depth data could be enabled in main.py)

• --out_dir - sets output directory

More details can be found in main.py and by using

python main.py --help

Surface normals can be visualised by editing renderoptions.json

2. standard reconstruction

Remove --d 0 from command above:

  python main.py --dataset kitti_0003 --r --c --f

3. debug

Inspect (trajetory-ID colour coded) 2D bboxes

python main.py --dataset kitti_0003 --show_2Dbboxes

Similarly, we can inspect 3D bboxes using --show_3Dbboxes or combine both arguments.

====

Note this is an un-optimised research code which should be cleaned up at some point (pull requests welcomed!)