Reconstruction of cue-invariant 3D appearances from brain activity

Setup

git submodule update --init --recursive

Steps

1. Feature extraction — Extract DNN features from 3D point clouds using AtlasNet. See feature-extraction/atlasnet/README.md for instructions.

2. Feature decoding — Decode DNN features from fMRI data using linear decoders. This includes cross-validation training and prediction for estimating the decoded-feature scaling factor. See feature-decoding/README.md for instructions.

3. Feature scaling and reconstruction — Scale decoded features using the cross-validation estimate, then reconstruct 3D shapes from true or scaled decoded features using AtlasNet. See reconstruction/atlasnet/README.md for instructions.