This repository contains the project created for my Masters thesis Neural 3D Reconstruction and Stylization.
Using IDR, we create a 3D reconstruction which we then style using NNFM from ARF or the original NST style approach.
The code was run using python 3.9, and requires packages which can be installed by running
pip install -r requirements.txt
Create a 3D reconstruction using IDR of the BK dataset by running
cd opt
python run_IDR.py --conf DTU_style.conf --scan_id 105
Once this has been completed the result is stored in the created exp folder. The result can be styled by running
cd opt
python run_style.py --conf DTU_style.conf --scan_id 24 --style vangogh_starry_night
Dataset from IDR.
We have also created a dataset of the TU Delft Architecture building, included in the repository.
To create your own dataset, camera information is required which can be calculated using COLMAP. As input the following is required
- <data_folder>
- images # input images
- masks # mask data of the input images
Run COLMAP using the following command-lines
colmap feature_extractor \
--database_path <data_folder>/database.db \
--image_path <data_folder>/image
colmap exhaustive_matcher \
--database_path <data_folder>/database.db
mkdir <data_folder>/sparse
colmap mapper \
--database_path <data_folder>/database.db \
--image_path <data_folder>/image \
--output_path <data_folder>/sparse
colmap model_converter \
--input_path <data_folder>/sparse/0 \
--output_path <data_folder>/sparse \
--output_type TXT
To create the required camera paremeters as well as normalized camera parameters (see IDR), run the following
python colmap2idr.py --dense_folder <data_folder> --max_d 256 --convert_format
python3 preprocess_cameras.py --source_dir <data_folder>