Home

Awesome

TRoVE

This repository provides code for the paper TRoVE accepted to ECCV 2022.

Requirements

Please make sure the following are installed and setup correctly to run the code provided in this repository correctly.

Python dependencies

  1. Setup blenderproc for Blender 3.0 and Python 3.6.9 (Not tested for different versions yet)
  2. skimage, tqdm, pyquaternion, opencv-python, h5py
  3. nuscenes-devkit (for using nuscenes dataset)

Assets (3D Objects and Textures)

The 3D objects used in the paper can be downloaded at 3D Objects (drive) and saved in a directory 3d_objects, and the texture data from Textures (drive) and saved in PBR_textures.

nuScenes setup

To run the code and experiments from the paper, the nuscenes dataset annotations are required only. The actual images and LiDAR/Radar is optional. The dataset annotations shall be downloaded to a root folder in the nuscenes dataset format and the path will be used in the configuration of the scripts.

Blender setup

Blender 3.0 was used in the development of this project using Blenderproc and additional addons such as blender-osm.

Data generation

Make sure to check each script for configurations and setup data paths. First, we prepare to parse the data from nuscenes and prepare the scene meta-data, and then after that we want to generate the OSM data needed for map preparation. For this, run the follwing two scripts:

python generate_nuscenes_config.py

python generate_osm_map_script.py

Once the data is prepared, we shall have 4 main folders: nuscenes config, osm data, 3d_objects, and PBR_textures. Once this is available, prepare blender and the addons by running:

blenderproc run configure_blender_addons.py

Blenderproc will download a version of blender if one is not available or set already. After addon setup, update the asset configurations (especially perform this step is paths or machines have been changed)

blenderproc run update_3d_asset_config.py

And finally, to generate data from a scene sample, run the following command:

blenderproc run run_trove_once.py

The above command will generate data for the default scene. To generate data for another scene (say scene 5, sample 7), run:

blenderproc run run_trove_once.py --scene_config ./config_dump/scene_0005/sample_007.json

The command will generate 24 outputs on each run. 6 images from the ego-vehicle in the dataset for the cameras and 18 from 3 random vehicles. The generated data is available in HDF5 format.