d_drand

---

Background Generation

First, create background images:

python vesselfm/d_drand/gen_bg.py

In our experiments, we created 50,000 background images. To be specific, we created 20,000 background images with no background geometry and perlin noise (mode_bg_geom=None, mode_bg_noise='perlin'), 10,000 with polyhedrons and perlin noise (mode_bg_geom='voronoi, mode_bg_noise='perlin'), 10,000 with spheres and perlin noise (mode_bg_geom='spheres, mode_bg_noise='perlin'), and 10,000 plain images (mode_bg_geom=None, mode_bg_noise='plain'). Please set the out_dir and adjust gen_bg.py as desired (see # TODO).

Data Generation

Finally, to create our domain randomized dataset $\mathcal{D}_\text{drand}$, run:

python vesselfm/d_drand/gen_data.py

Note, to first set OUT_DIR (path to output dir), DATASET_LABEL (path to foreground geometry), and DATASET_BACKGROUND (path to previously generated background images) in the config file (see #TODO). $\mathcal{D}_\text{drand}$ will be stored in OUT_DIR and contain 500,000 image-mask pairs.

The final structure of $\mathcal{D}_\text{drand}$ is as follows:

/path/to/d_drand   # D_drand
└── 0/   # sample 0
    └── img.npy   # domain randomized image of shape 128x128x128
    └── mask.npy  # matching label of shape 128x128x128
└── 1/
    └── img.npy 
    └── mask.npy
...

└── 499999/
    └── img.npy
    └── mask.npy