VPP: Efficient Conditional 3D Generation via Voxel-Point Progressive Representation

Authors: Zekun Qi, Muzhou Yu, Runpei Dong, Kaisheng Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that VPP efficiently generates high-fidelity and diverse 3D shapes across different categories, while also exhibiting excellent representation transfer performance.
Researcher Affiliation Academia Zekun Qi Muzhou Yu Runpei Dong Xi an Jiaotong University {qzk139318, muzhou9999, runpei.dong}@stu.xjtu.edu.cn Kaisheng Ma Tsinghua University kaisheng@mail.tsinghua.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. Figure 11 is a diagram/flowchart, not pseudocode.
Open Source Code Yes Codes will be released at https://github.com/qizekun/VPP.
Open Datasets Yes We use Shape Net Core from Shape Net [4] as the training dataset. ... Scan Object NN [83] and Model Net [92] are currently the two most challenging 3D object datasets...
Dataset Splits No The paper mentions using Shape Net for training and Scan Object NN/Model Net40 for transfer classification, but it does not specify explicit training, validation, and test split percentages or sample counts in the main text or appendix.
Hardware Specification Yes Notably, VPP is capable of generating high-quality 8K point clouds within 0.2 seconds on a single RTX 2080Ti. ... Table 4: GPU device NVIDIA A100.
Software Dependencies No The paper mentions software components like "Optimizer Adam Adam W" and "Learning rate scheduler cosine" in Table 4. However, it does not provide specific version numbers for these software components or the underlying frameworks (e.g., PyTorch, TensorFlow) used for implementation.
Experiment Setup Yes Table 4: Training recipes for 3D VQGAN, Voxel Generator, Grid Smoother and Point Upsampler. This table includes specific hyperparameters such as Learning rate (1e-4, 1e-3), Weight decay (1e-4, 5e-2), Training epochs (100, 300), Warmup epochs (5, 10), Batch size (32, 128), and Drop path rate (0.1).