A PyTorch 0.4+ implementation of the proof-of-concept experiment from Implicit 3D Orientation Learning for 6D Object Detection from RGB Images by Martin Sundermeyer et al. ECCV 2018.
git clone https://github.com/lwneal/implicit3d
cd implicit3d
pip install -r requirements.txt
python denoising_autoencoder.pyPosition and scale are treated as noise factors, and the autoencoder learns to be invariant to them, learning only the desired factor (rotation).
Graph showing the value of the encoding of an image as the rotation of the image changes.
Note the period of the graph- due to rotational symmetry, the same representation is repeated with four offsets.
Top 4: Original images. Bottom 4: Reconstructed images after training.