Template-free Articulated Neural Point Clouds for Reposable View Synthesis

Authors: Lukas Uzolas, Elmar Eisemann, Petr Kellnhofer

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our work... Datasets We chose three multi-view video datasets... We compare our method to state-of-the-art non-articulated and articulated methods... Novel view synthesis We provide results for the Robots view synthesis without the skeleton simplification; quantitatively in Fig. 4 and qualitatively in Fig. 3. Loss ablation Our experiments suggest the regularization does not improve the image quality, but it improves the quality of the kinematic model, which is important for our main goal of reposing (see Fig. 9).
Researcher Affiliation Academia Lukas Uzolas Elmar Eisemann Petr Kellnhofer Delft University of Technology The Netherlands {l.uzolas, e.eisemann, p.kellnhofer}@tudelft.nl
Pseudocode No No explicit pseudocode or algorithm blocks were found in the paper.
Open Source Code No The project website can be found at https://lukas.uzolas.com/Articulated-Point-Ne RF/. This link points to a project website/demonstration page, not an explicit code repository or statement about open-source code availability for the described methodology.
Open Datasets Yes Datasets We chose three multi-view video datasets that are commonly used for the evaluation of dynamic multi-view synthesis methods and pose articulation. First, the Robots dataset [20]... Second, the Blender dataset [3]... Third, ZJU-Mo Cap dataset [82]...
Dataset Splits Yes We use 18 views for training and 2 for evaluation. We use the original training-test split. We use the original training-test split. We use the original training-test split. We use the original training-test split. Finally, we evaluate image quality using peak signal-to-noise ratio (PSNR), structural similarity (SSIM) [114], and learned perceptual image patch similarity (LPIPS) [115] image metrics.
Hardware Specification Yes All experiments were done on a single Nvidia GPU RTX 3090Ti.
Software Dependencies No The paper mentions "implemented in Pytorch" and using "Ti Neu Vox [11] backbone using the authors implementation", but no specific version numbers for these software dependencies are provided.
Experiment Setup Yes We train each scene using the Adam optimizer for 160k (Blender and Robots) or 320k (ZJU-Mo Cap) iterations with a batch size of 8192 rays, sampled randomly from multiple views and a single timestamp. In total, our training loss is L = ω0Lphoto + ω1Lmask + ω2Lskel + ω3Ltranf + ω4Lsmooth + ω5Lsparse + ω6LARAP where ω = {200, 0.02, 1, 0.1, 10, 0.2, 0.005} in our experiments.