Directed Distance Function

This repository contains all the code needed to train a network to learn the directed distance to an object surface. This includes code for data generation, model creation, training, and visualization.

Requirements

• trimesh
• tqdm
• scikit-learn
• matplotlib
• pytorch
• open3d
• pickle (part of standard Python)
• palettable
• rtree
• igl

More requirements for Srinath's v2:

• beacon (pip install git+https://github.com/brown-ivl/beacon.git)
• tk3dv (pip install git+https://github.com/drsrinathsridhar/tk3dv.git)
• [ Required on some machines to avoid an OpenMP issue ] conda install -c conda-forge nomkl

Data Generation / Sampling Rays

The data generation code offers a few methods for sampling rays with the idea that some sampling techniques produce harder to learn rays (edge cases). To visualize the different sampling methods, use

python sampling.py -v --mesh_file <path to .obj --use_4d>

You can also see how fast each sampling method is by running

python sampling.py -s

Training

To train, test, and save a network, run

python train4D.py -Tts -n mynetwork --mesh_file <path to .obj> --save_dir <dir results can be written in> --intersect_limit <# of intersections>

The flags -d, -v, -p, and -m can be used to create depth images, depth video, point clouds, and meshes respectively. The video will be saved to <save_dir>/depth_videos while the rest of the visualizations will be displayed on screen.

To see all flags use

python train4D.py --help

Codebase Overview

• Training/Testing - train4D.py
• Data Generation - data.py, sampling.py
• Network - model.py
• Utility Functions - utils.py, rasterization.py
• Visualization - camera.py, visualization.py