Awesome

This is the official code of Few-Shot Generalization for Single-Image 3D Reconstruction via Priors (ICCV 2019).

ShapeNet rendered images and voxels as used in this work are available for download at the R2N2 Github. Once downloaded, change the paths in L14/15 of highres_sampler.py.

The main training script is train_iterative_RGB_refiner.py. The most important argument is --excluded-cats which dictates which categories to hold out for few-shot learning. Most experiments presented in our paper have --excluded-cats benches,cabinets,lamps,sofas,vessels,rifles.

Other scripts in extra_files/ are remnants of various explorations/evaluations many of which didn't make it into the main paper. Them working out of the box is relatively improbable, but feel free to experiment with these and let me know if you find anything interesting!

Also please let me know if you have any questions concerning this code or the paper itself.