DataLoaders

This contains various data loaders for a series of datasets:

• DeepVoxels : Synthetic images and depth of objects captured in a dome
• NeRF : Synthetic images of objects captured in a dome
• LLFF : Real images of objects captures in a front-facing manner around a semi-dome
• ScanNet : Point clouds of rooms with annotations for 20 classes (bed, furnite, wall, etc.)
• SemanticKitti : Point clouds from a car driving in urban scenario annotated with 19 classes.
• ShapeNetSem : Point clouds of various objects (airplance, motorbike) with part-based annotations
• VolumetricRefinement : RGB-D images used in the work of Zollhöfer et al: Shading-based Refinement on Volumetric Signed Distance Functions
• Pheno4D : Point clouds of maize and tomato plants together with instance segmentation annotations

Example of usage and loading of the data can be found in python/test_loader.py Each loader is controlled by a config file which can be found in config/test_loader.cfg

Build and install:

To build and install the example, you must have first installed EasyPBR. Afterwards the dataloader can be cloned and compiled with

$ git clone --recursive https://github.com/RaduAlexandru/data_loaders
$ cd data_loaders
$ make

Optionally, if you have Robot Operating System (ROS), you can clone data_loaders in your catkin workspace and you will have access to dataloaders for ROS bags and ROS topics.

$ cd YOUR_ROS_WORKSPACE/src
$ git clone --recursive https://github.com/RaduAlexandru/data_loaders
$ cd data_loaders
$ make

Usage:

There are examples on how to use each data loader in ./python/test_loader.py and the file ./config/test_loader.cfg describes the parameters to configure each dataset.

Links:

• DeepVoxels :
  • https://drive.google.com/uc?id=1lUvJWB6oFtT8EQ_NzBrXnmi25BufxRfl
  • It was used in IBRNet and the link for the google drive is from https://github.com/googleinterns/IBRNet/blob/master/data/download_eval_data.sh
• VolumetricRefinement :
  • From the paper Shading-based Refinement on Volumetric Signed Distance Functions http://graphics.stanford.edu/projects/vsfs/#data
• Nerf :
  • https://drive.google.com/uc?id=18JxhpWD-4ZmuFKLzKlAw-w5PpzZxXOcG
  • It was used in IBRNet and the link for the google drive is from https://github.com/googleinterns/IBRNet/blob/master/data/download_eval_data.sh
• LLFF :
  • https://drive.google.com/uc?id=16VnMcF1KJYxN9QId6TClMsZRahHNMW5g
  • It was used in IBRNet and the link for the google drive is from https://github.com/googleinterns/IBRNet/blob/master/data/download_eval_data.sh

Misc:

This code uses an "OpenCV" style camera coordinate system, where the Y-axis points downwards (the up-vector points in the negative Y-direction), the X-axis points right, and the Z-axis points into the image plane.