Neural 3D Reconstruction in the Wild
Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, Noah Snavely
SIGGRAPH 2022 (Conference Proceedings)

• Training (i.e., reconstruction) code.
• Toolkit and pipeline to reproduce the evaluation results on the proposed Heritage-Recon dataset.
• Config for reconstructing generic outdoor/indoor scenes.

conda env create -f environment.yaml
conda activate neuconw
scripts/download_sem_model.sh

Download the Heritage-Recon dataset and put it under data. You can also use gdown to download it in command line:

mkdir data && cd data
gdown --id 1eZvmk4GQkrRKUNZpagZEIY_z8Lsdw94v

Generate ray cache for all four scenes:

for SCENE_NAME in brandenburg_gate lincoln_memorial palacio_de_bellas_artes pantheon_exterior; do
  scripts/data_generation.sh data/heritage-recon/${SCENE_NAME}
done

To train scenes in our Heritage-Recon dataset:

# Subsutitude `SCENE_NAME` with the scene you want to reconstruct.
scripts/train.sh $EXP_NAME config/train_${SCENE_NAME}.yaml $NUM_GPU $NUM_NODE

First, extracting mesh from a checkpoint you want to evaluate:

scripts/sdf_extract.sh $EXP_NAME config/train_${SCENE_NAME}.yaml $CKPT_PATH 10

The reconstructed meshes will be saved to PROJECT_PATH/results.

Then run the evaluation pipeline:

scripts/eval_pipeline.sh $SCENE_NAME $MESH_PATH

Evaluation results will be saved in the same folder with the evaluated mesh.

The code takes a standard COLMAP workspace format as input, a script is provided for automatically convert a colmap workspace into our data format:

scripts/preprocess_data.sh

More instructions can be found in scripts/preprocess_data.sh

However, if you wish to select a better bounding box (i.e., reconstruction region) manually, do the following steps.

Generate semantic maps:

python tools/prepare_data/prepare_semantic_maps.py --root_dir $WORKSPACE_PATH --gpu 0

Create a file config.yaml into workspace to write metadata. The target scene needs to be normalized into a unit sphere, which require manual selection. One simple way is to use SFM key-points points from COLMAP to determine the origin and radius. Also, a bounding box is required, which can be set to [origin-raidus, origin+radius], or only the region you're interested in.

{
    name: brandenburg_gate, # scene name
    origin: [ 0.568699, -0.0935532, 6.28958 ], 
    radius: 4.6,
    eval_bbx: [[-14.95992661, -1.97035599, -16.59869957],[48.60944366, 30.66258621, 12.81980324]],
    voxel_size: 0.25,
    min_track_length: 10,
    # The following configuration is only used in evaluation, can be ignored for your own scene
    sfm2gt: [[1, 0, 0, 0],
            [ 0, 1, 0, 0],
            [ 0, 0, 1, 0],
            [ 0, 0, 0, 1]],
}

Run the following command with a WORKSPACE_PATH specified:

scripts/data_generation.sh $WORKSPACE_PATH

After completing above steps, whether automatically or manually, the COLMAP workspace should be looking like this;

└── brandenburg_gate
  └── brandenburg_gate.tsv
  ├── cache_sgs
    └── splits
        ├── rays1_meta_info.json
        ├── rgbs1_meta_info.json
        ├── split_0
            ├── rays1.h5
            └── rgbs1.h5
        ├── split_1
        ├──.....
  ├── config.yaml
  ├── dense
    └── sparse
        ├── cameras.bin
        ├── images.bin
        ├── points3D.bin
  └── semantic_maps
      ├── 99119670_397881696.jpg
      ├── 99128562_6434086647.jpg
      ├── 99250931_9123849334.jpg
      ├── 99388860_2887395078.jpg
      ├──.....

Change DATASET.ROOT_DIR to COLMAP workspace path in config/train.yaml, and run:

scripts/train.sh $EXP_NAME config/train.yaml $NUM_GPU $NUM_NODE

Additionally, NEUCONW.SDF_CONFIG.inside_outside should be set to True if training an indoor scene (refer to config/train_indoor.yaml).

scripts/sdf_extract.sh $EXP_NAME config/train.yaml $CKPT_PATH $EVAL_LEVEL

The reconstructed meshes will be saved to PROJECT_PATH/results.

If you find this code useful for your research, please use the following BibTeX entry.

@inproceedings{sun2022neuconw,
  title={Neural {3D} Reconstruction in the Wild},
  author={Sun, Jiaming and Chen, Xi and Wang, Qianqian and Li, Zhengqi and Averbuch-Elor, Hadar and Zhou, Xiaowei and Snavely, Noah},
  booktitle={SIGGRAPH Conference Proceedings},
  year={2022}
}

Part of our code is derived from nerf_pl and NeuS, thanks to their authors for the great works.