Light Field Networks: Neural Scene Representations with Single-Evaluation Rendering

Authors: Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, Fredo Durand

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of LFNs by reconstructing 360-degree light fields of a variety of simple 3D scenes. In all experiments, we parameterize LFNs via a 6-layer Re LU MLP, and the hypernetwork as a 3-layer Re LU MLP, both with layer normalization. We solve all optimization problems using the ADAM solver with a step size of 10 4. Please find more results, as well as precise hyperparameter, implementation, and dataset details, in the supplemental document and video. Table 1: Single-shot multi-class reconstruction results. We benchmark LFNs with SRNs [3] and DVR [5] on the task of single-shot (auto-decoding with a single view), multi-class reconstruction of the 13 largest Shape Net [59] classes.
Researcher Affiliation Academia Vincent Sitzmann1, sitzmann@mit.edu Semon Rezchikov2, skr@math.columbia.edu William T. Freeman1,3 billf@mit.edu Joshua B. Tenenbaum1,4,5 jbt@mit.edu Frédo Durand1 fredo@mit.edu 1MIT CSAIL 2Columbia University 3 IAFI 4MIT BCS 5 CBMM
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions 'please see the code for implementation details' and refers to a project page 'vsitzmann.github.io/lfns/' and 'supplemental document and video', but does not provide a concrete access link to a source code repository or an explicit statement of code release for the methodology described within the main paper.
Open Datasets Yes We train LFNs on the Shape Net cars dataset with 50 observations per object from [3], as well as on simple room-scale environments as proposed in [13]. ShapeNet: An information-rich 3d model repository. ar Xiv preprint ar Xiv:1512.03012, 2015.
Dataset Splits Yes Multi-class single-view reconstruction. Following [5, 6], we benchmark LFNs with recent global conditioning methods on the task of single-view reconstruction and novel view synthesis of the 13 largest Shape Net categories. We follow the same evaluation protocol as [60] and train a single model across all categories.
Hardware Specification Yes All clock times were collected for rendering 256 256 images on an NVIDIA RTX 6000 GPU.
Software Dependencies No The paper mentions using the 'ADAM solver' but does not provide specific software names with version numbers for libraries or frameworks used in the experiments.
Experiment Setup Yes In all experiments, we parameterize LFNs via a 6-layer Re LU MLP, and the hypernetwork as a 3-layer Re LU MLP, both with layer normalization. We solve all optimization problems using the ADAM solver with a step size of 10 4.