Fast Training of Neural Lumigraph Representations using Meta Learning

Authors: Alexander Bergman, Petr Kellnhofer, Gordon Wetzstein

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show that Meta NLR++ achieves similar or better novel view synthesis results in a fraction of the time that competing methods require. ... We demonstrate training of high-quality neural scene representations in minutes or tens of minutes, rather than hours or days, which can then be rendered at real-time framerates. ... We demonstrate this by comparison to several state-of-the-art methods. Specifically, we evaluate the volumetric representation of Ne RF [60], a meshbased representation similar to SVS [8], the neural signed distance function-based representations of IDR [4] and NLR [5], and the image-based rendering of IBRNet [29].
Researcher Affiliation Academia Alexander W. Bergman Stanford University awb@stanford.edu Petr Kellnhofer Stanford University pkellnho@stanford.edu Gordon Wetzstein Stanford University gordon.wetzstein@stanford.edu
Pseudocode No The paper does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The source code and pre-trained models are available on our project website, and the full set of implementation details including hyperparameters, training schedules, and architectures are described in our supplement for each of the various datasets evaluated on. computationalimaging.org/publications/metanlr/
Open Datasets Yes For the following comparisons, we use the DTU dataset [4, 88], which has been made public by its creators... To show that Meta NLR++ is robust to datasets beyond DTU, we evaluate using the multi-view dataset in NLR [5]. ... Additional results demonstrating this capability on the Shape Net [90] dataset are included in the supplementary document.
Dataset Splits No No explicit percentages or absolute counts for training, validation, and test splits are provided in the main text. The paper mentions using 'seven ground-truth views' for training on DTU and 'three held-out test views', but does not detail a formal validation split or clear splitting methodology for reproducibility.
Hardware Specification Yes We train each of our models on a single Nvidia Quadro RTX8000 GPU. We also use an Nvidia Quadro RTX6000 GPU for rendering and training iteration time computation. In total, we have an internal server system with four Nvidia Quadro RTX8000 GPUs and six Nvidia Quadro RTX6000 GPUs, of which we used a subset of three RTX8000s and one RTX6000.
Software Dependencies No The paper states 'We implement Meta NLR++ in Py Torch and use the Adam [87] optimizer for all optimization steps', but it does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We implement Meta NLR++ in Py Torch and use the Adam [87] optimizer for all optimization steps of NLR++, with a starting learning rate of 1 10 4 for Φ and 5 10 4 for E, D, Γ. We use α = 50, τ = 1 10 3, and β = 1 10 1 as starting hyperparameter values, which are progressively decayed (or increased in the case of α) through training (full schedules are described in the supplement). We use shape loss training hyperparameter values of t1 = 50 and t2 = 7, and loss weight parameters of λ1 = 1 102/α, λ2 = 3.