Awesome

FineRecon

This software project accompanies the research paper, FineRecon: Depth-aware Feed-forward Network for Detailed 3D Reconstruction.

FineRecon is a deep learning model for 3D reconstruction from posed RGB images.

Setup

Dependencies

pip install \
  matplotlib \
  pillow \
  numpy \
  scikit-image \
  scipy \
  timm \
  torch==1.13 \
  torchvision \
  "tqdm>=4.65" \
  trimesh \
  pytorch_lightning==1.8 \
  pyyaml \
  opencv-python-headless \
  python-box \
  tensorboard

Config

cp example-config.yml config.yml

The paths in config.yml will need to be edited to point to the data directories.

Data

FineRecon requires an RGB-D scan dataset such as ScanNet, which can be downloaded and extracted using the scripts provided by the ScanNet authors.

The dataset structure expected by FineRecon is

/path/to/dataset/
    test.txt
    train.txt
    val.txt
    first_scan/
        color/
            0.jpg
            1.jpg
            2.jpg
            ...
        depth/
            0.png
            1.png
            2.png
            ...
        intrinsic_color.txt
        intrinsic_depth.txt
        pose.npy
    second_scan/
    ...
    last_scan/

The files test.txt, train.txt, and val.txt should each contain a newline-separated list of scan directory names (e.g. first_scan) describing the test, train, and validation splits respectively. Each pose.npy contains the camera poses (world-to-camera transformation matrices) as an array of size (N, 4, 4) in npy format, where any invalid poses are marked with the value Inf. The files intrinsic_color.txt and intrinsic_depth.txt should contain the (4, 4) color and depth intrinsic matrices, respectively. In config.yml, the value of dataset_dir should be set to /path/to/dataset.

To generate the ground truth TSDF run generate_gt_tsdf.py --dataset-dir /path/to/dataset --output-dir /path/to/gt_tsdf, and in config.yml set the value of tsdf_dir to /path/to/gt_tsdf. To run training or inference with depth guidance, make sure depth_guidance.enabled is set to True in the config and set the value of depth_guidance.pred_depth_dir to /path/to/pred_depth, which should have the following structure:

/path/to/pred_depth/
    first_scan/
        depth/
            0.png
            1.png
            2.png
            ...
        intrinsic_depth.txt
    second_scan/
    ...
    last_scan/

It can be helpful to limit inference to only using a set of pre-defined keyframes, because it's faster (particulary with point back-projection enabled) and because depth estimates may not be available for all frames. To do this, set test_keyframes_file in the config to the location of a JSON file with the following structure:

{
  "first_scan": [i0, i1, i2, ...],
  ...
}

where i0, i1, etc. are the integer indices of the keyframes.

Training

python main.py

Inference

We provide pre-trained weights here: checkpoint.zip. These are weights for our main model using resolution-agnostic TSDF supervision, depth guidance, and point-backprojection.

python main.py --task predict --ckpt path/to/checkpoint.ckpt

For convenience, we also provide the inference results (meshes) of our main model on the ScanNet test set:

• High-resolution [1 cm] (2.4 GB) → This is the resolution used in figures and metrics, unless otherwise stated.
• Low-resolution [4 cm] (148 MB)

Evaluation

Evaluation code and data for 3D metrics can be found in TransformerFusion, and evaluation code for 2D metrics can be found in Atlas.

Citation

@article{stier2023finerecon,
  title={{FineRecon}: Depth-aware Feed-forward Network for Detailed 3D Reconstruction},
  author={Stier, Noah and Ranjan, Anurag and Colburn, Alex and Yan, Yajie and Yang, Liang and Ma, Fangchang and Angles, Baptiste},
  journal={arXiv preprint},
  year={2023}
}