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

Authors: Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, Matthew O'Toole

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments, Table 2: Phasor supervision aids few-view reconstruction. Each cell contains RGB image similarity measures, and metrics are computed on 10 hold-out views.
Researcher Affiliation Collaboration Benjamin Attal Carnegie Mellon University Eliot Laidlaw Brown University Aaron Gokaslan Cornell University Changil Kim Facebook Christian Richardt University of Bath James Tompkin Brown University Matthew O Toole Carnegie Mellon University
Pseudocode No The paper describes methods through text and mathematical equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper links to a project website (imaging.cs.cmu.edu/torf) and mentions videos, but does not explicitly state that the source code for the methodology described is available there or elsewhere.
Open Datasets No The paper states, 'We captured the Phone Booth, Cup, Photocopier, Desk Box, and Study Book sequences with our handheld camera setup.' and 'Finally, we create synthetic raw C-To F sequences Bathroom, Bedroom, and Dino Pear by adapting the physically-based path tracer PBRT [40] to generate phasor images with multi-bounce and scattering effects.' These are custom datasets, and no access information (link, DOI, or explicit public availability) is provided.
Dataset Splits No The paper refers to 'hold-out views' used for evaluation, which serve as test sets, but it does not explicitly describe a distinct validation set or its specific split details for hyperparameter tuning during training.
Hardware Specification Yes For optimization, we use an NVIDIA Ge Force RTX 2080 Ti with 11 GB RAM.
Software Dependencies No The paper mentions 'We use Open CV to calibrate intrinsics, extrinsics, and lens distortion.' but does not specify a version number for OpenCV or any other software dependencies.
Experiment Setup Yes First, we optimize the weights of the static neural network F stat θ , as well as the camera poses for each video frame and the relative rotation and translation between the color and C-To F sensor, with a learning rate of 10 3. After 5000 iterations, we decrease the pose learning rate to 5 10 4, and optimize our full model. At training time, we reduce the weight λ in later iterations to prioritize the color loss (halved every 125,000 iterations).