ProvNeRF: Modeling per Point Provenance in NeRFs as a Stochastic Field
Authors: Kiyohiro Nakayama, Mikaela Angelina Uy, Yang You, Ke Li, Leonidas J. Guibas
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that modeling per-point provenance during the Ne RF optimization enriches the model with information on triangulation leading to improvements in novel view synthesis and uncertainty estimation under the challenging sparse, unconstrained view setting against competitive baselines1. 5 Experiments Our Prov Ne RF learns per-point provenance field Dθ by optimizing LProv Ne RF on a Ne RF-based model. To validate Prov Ne RF, we demonstrate that jointly optimizing the provenance distribution and Ne RF representation can result in better scene reconstruction as shown in the task of novel view synthesis (Sec. 5.1). Moreover, we also show that the learned provenance distribution enables other downstream tasks such as estimating the uncertainty of the capturing field (Sec. 5.2). We provide an ablation study on f IMLE against other probabilistic methods in Sec. 5.3. |
| Researcher Affiliation | Collaboration | Kiyohiro Nakayama1 Mikaela Angelina Uy1,2 Yang You1 Ke Li3 Leonida J. Guibas1 1 Stanford University 2 Nvidia 3 Simon Fraser University |
| Pseudocode | No | The paper provides a training pipeline illustration (Figure 3) but does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | No | 1Code will be available at https://github.com/george Nakayama/Prov Ne RF. Justification: We will provide reproducible code upon the paper s acceptance. |
| Open Datasets | Yes | Table 1: Novel View Synthesis Results. Our method outperforms baselines in novel view synthesis on both Scannet and Tanks and Temple Datasets. Scannet [13] and Matterport3D [9]. |
| Dataset Splits | No | The paper refers to supplementary material for dataset details ('See the supplement for details on the dataset, metrics, baselines, and implementation details') but does not specify explicit training/validation/test splits in the main text. |
| Hardware Specification | Yes | The post-training takes around 30 minutes on a single A6000 Nvidia GPU. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks used in the experiments. |
| Experiment Setup | No | The paper states that 'implementation and architectural details' and 'training details, hyperparameters, network architectures' are provided in the supplementary material, but it does not include specific experimental setup details or hyperparameter values in the main text. |