In this work, we are focusing on reconstructing scenes from a single image. We make use of IF-Net which focuses on shape completion from an incomplete 3D input. We explore the usage of IF-Net in the task of 3D reconstruction from images. We also explore its ability to work on complex scenes instead of simple shapes as was proposed in the paper. For more details, please check the full report and presentation of the project.
For a brief 'Real-World-Experiment' see the test-report of the proposed architecture, note however that we used a more powerful depth regressor here real-world-test.
A linux system with cuda 9.0 is required.
Install the dependencies with conda using the 3d-recon_env.yml file :
conda env create -f 3d-recon_env.yml
conda activate 3d-recon
Clone the repository and navigate into it in the terminal.
Install libmesh needed libraries with:
cd data_processing/libmesh/
python setup.py build_ext --inplace
cd ../..
In this work, we used 3D-Front dataset.
- IF-Nets by [Chibane et. al. 2020]
- SynSin by [Olivia Wiles et al. 2019]
- point clouds by puhsu
- Occupancy Networks by [Mescheder et. al. CVPR'19]
- DISN by [Xu et. al. NeurIPS'19]
- IF-Nets by [Chibane et. al. 2020]
- Unsupervised learning of shape and pose with differentiable point clouds by [Eldar Insafutdinov and Alexey Dosovitskiy 2018]
- SynSin: End-to-end View Synthesis from a Single Image by [Olivia Wiles et al. 2019]