Monocular Dynamic View Synthesis: A Reality Check

Authors: Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, Angjoo Kanazawa

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive evaluation on existing datasets [5, 7] as well as a new dataset that includes more challenging motion and diverse scenes. In this section, we conduct a series of empirical studies to disentangle the recent progress in dynamic view synthesis (DVS) given a monocular video from effective multi-view in the training data.
Researcher Affiliation Collaboration Hang Gao1, Ruilong Li1 Shubham Tulsiani2 Bryan Russell3 Angjoo Kanazawa1 1UC Berkeley 2Carnegie Mellon University 3Adobe Research
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code and data can be found at http://hangg7.com/dycheck.
Open Datasets Yes We investigate the datasets used for evaluation in D-Ne RF [3], Hyper Ne RF [7], Nerfies [5], and NSFF [4]. Table 1 shows their statistics. We evaluate the amount of effective multi-view cues via the proposed EMFs, shown in Figure 3. We find that existing datasets have large EMF values on both metrics. ... We also propose a new dataset called the i Phone dataset shown in Figure 7.
Dataset Splits No We train all approaches under these two settings with the same held-out validation frames and same set of co-visibility masks computed from common training frames. (This indicates validation is used, but not the specific split sizes or percentages).
Hardware Specification No The paper does not explicitly mention specific hardware details such as GPU/CPU models, memory, or cloud instance types used for experiments.
Software Dependencies No We consolidate Nerfies [5] and Hyper Ne RF [7] in one codebase using JAX [46]. (The reference [46] points to 'JAX: composable transformations of Python+Num Py programs, 2018.' which gives a year but not a specific version number of the software itself.)
Experiment Setup Yes We consolidate Nerfies [5] and Hyper Ne RF [7] in one codebase using JAX [46]. Compared to the original official code releases, our implementation aligns all training and evaluation details between models and allows correspondence readout. Concretely, we consider the following: (+B) random background compositing [32]; (+D) a depth loss on the ray matching distance [1, 4]; and (+S) a sparsity regularization for scene surface [49].