Interactive Reconstruction of Monte Carlo Image Sequences using a Recurrent Denoising Autoencoder
Tensorflow Implementation and some visual results of SIGGRAPH'17 paper: Interactive Reconstruction of Monte Carlo Image Sequences using a Recurrent Denoising Autoencoder by Chakravarty R. Alla Chaitanya, Anton Kaplanyan, Christoph Schied, Marco Salvi, Aaron Lefohn, Derek Nowrouzezahrai and Timo Aila
Architecture
We use the same architecture proposed in the paper instead we change the recurrent block to ConvLSTM blocks.
Training
python main.py --dataPath /path/to/data --outputPath /path/to/output_folder Data Acquisition
We use pbrt-v3 to render 1 spp, 4096 spp, and auxiliary features as training pairs. Example input features in grey scale and reference in bathroom scene (a) 1 spp MC rendering with Halton sampler, (b) depth, (c) shading normal x-axis, (d) shading normal y-axis, (e) roughness, and (f) 4096 spp reference.
Example of our training sequence:
Visual Results
Right: Noisy 1 spp input RGB sequence
Middle: Reconstructed sequence
Left: Reference 4096 spp sequence
