TöRF: Time-of-Flight Radiance Fields for Dynamic Scene View Synthesis

Environment setup

Download data from here and place its contents in the ./data folder

To set up a conda environment, run:

conda env create -f environment.yml
conda activate torf

Training

Quick start

To train model on one of our sequences, run

./scripts/real.sh seq_color21_small 0 30

The last two arguments specify first frame to use and the last frame to use

Dynamic sequences

For the real and iOS sequences, run

./scripts/real.sh seq_color21_small 0 30

and

./scripts/ios.sh dishwasher 30

respectively.

Static sequences

For the static sequences run

./scripts/static.sh bathroom_static 2 <width> <height> 0,8

Where the last argument specifies which views to use for training. Image dimensions are 512x512 for the bathroom sequence, and 640x360 for bedroom.

Using Your Own Data

Format

If your data contains ground truth ToF images, then you should use the RealDataset loader, or some variant of it, and the data should be formatted as follows.

cams/
    tof_intrinsics[.npy|.mat]
    tof_extrinsics[.npy|.mat]
    color_intrinsics[.npy|.mat]
    color_extrinsics[.npy|.mat]
color/
    [*.npy|*.mat]
tof/
    [*.npy|*.mat]

Note that each ToF image should contain (real, imaginary) components of the measured ToF phasor

Additional inputs

You can also optionally include:

cams/
    relative_R[.npy|.mat]
    relative_T[.npy|.mat]
    depth_range[.npy|.mat]
    phase_offset[.npy|.mat]

where relative_R and relative_T specify the relative pose of the color camera and ToF sensor. If you do not include these, they are initialized to identity. They can also be optimized during training with --use_relative_poses and --optimize_relative_pose

Depth

The file depth_range specifies a value that is twice unambiguous depth range of the ToF sensor, and phase_offset an offset between the measured phase and the true phase. The phase offset can also be optimized during training with --optimize_phase_offset

If metric depth, rather than ToF is available, then you can replace the tof folder with a depth folder. In this case, you should convert depth into ToF before training (see IOSDataset).

Extrinsics

We expect extrinsics (world to camera) in SfM format (y down, z forward). Note that the if the extrinsics are not scale correct (i.e. do not match the scale of depth / ToF), then they should be optimized during training with --optimize_poses

Additional flags

If your dataset contains collocated ToF / depth and color, you can add the flag --collocated_pose

Evaluation

Static evaluation

For evaluation on static sequences, run:

./scripts/static_eval.sh bathroom_static 2 <width> <height> 0,8

And then

python compute_metrics eval/[expname] 30

Citation

@article{attal2021torf,
  title={TöRF: Time-of-Flight Radiance Fields for Dynamic Scene View Synthesis},
  author={Attal, Benjamin and Laidlaw, Eliot and Gokaslan, Aaron and Kim, Changil and Richardt, Christian and Tompkin, James and O'Toole, Matthew},
  journal={Advances in neural information processing systems},
  volume={34},
  year={2021}
}

benattal/torf