gfxds 0.1.0

Dataloaders for multi-channel graphics datasets

Graphics Datasets in Rust (gfxds)

This Rust/Python package provides dataloaders for mult-channel image datasets, particularly related to material estimation and intrinsic image decomposition. Writing it in Rust was a little silly, but I've gotten so tired of working with torch.utils.data.DataLoader.

This library does as much as possible to make different datasets consistent, mapping all RGB color data as gamma=2.2 (not quite sRGB) and all non-color data as linear. Images should have reasonable exposure. Shading and albedo maps are computed so that image = albedo * shading / (1 - shading). Any flipped normals are corrected. Depth maps are rescaled (but not offset) to have a maximum of depth of 1. All pixel values are dequantized appropriately with random noise.

All default datasets can be found in datasets.toml. If you would like to add support for a new dataset, consider opening a PR with a new src/dataset/[whatever dataset].rs file and entry in the datasets.toml config. Regardless I would recommend forking this repo if you want to use it in your own research, since configuration is limited and most decisions are hard-coded.

Example

After downloading interiorverse (https://interiorverse.github.io/#download) to a datasets folder, you can use the dataloader like so:

from gfxds import Loader

loader = Loader('/path/to/the/datasets', 'train/interiorverse/85')
loader.start() # starts background worker threads

for sample in loader:
    sample.name # identifier for the element of the dataset
    sample.caption # text description
    for component, raster in sample.images.items():
        component # the name of the image (eg roughness, normals, image)
        raster # the usual float32 h*w*c numpy array

For the full API, see gfxds.pyi.