3DILG: Irregular Latent Grids for 3D Generative Modeling
Authors: Biao Zhang, Matthias Niessner, Peter Wonka
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show different applications that improve over the current state of the art. First, we show results for probabilistic shape reconstruction from a single higher resolution image. Second, we train a probabilistic model conditioned on very low resolution images. Third, we apply our model to category-conditioned generation. All probabilistic experiments confirm that we are able to generate detailed and high quality shapes to yield the new state of the art in generative 3D shape modeling. We set the size of the input point cloud to N = 2048. The number and the size of point patches are M = 512 and K = 32, respectively. We use the datset Shape Net-v2 [4] for shape reconstruction. We split samples into train/val/test (48597/1283/2592) set. Following the evaluation protocol of [13, 66], we include three metrics, volumetric Io U, the Chamfer-L1 distance, and F-Score [50]. |
| Researcher Affiliation | Academia | Biao Zhang KAUST biao.zhang@kaust.edu.sa Matthias Nießner Technical University of Munich niessner@tum.de Peter Wonka KAUST pwonka@gmail.com |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | 3. (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See Appendix. |
| Open Datasets | Yes | We use the datset Shape Net-v2 [4] for shape reconstruction. |
| Dataset Splits | Yes | We split samples into train/val/test (48597/1283/2592) set. |
| Hardware Specification | No | The paper only mentions '4-8 GPUs' in a general context of scalability. It does not provide specific hardware models (GPU/CPU) or detailed specifications used for running its experiments. The checklist points to the Appendix, which is not provided. |
| Software Dependencies | No | The paper mentions 'GPT [42]' but does not provide specific version numbers for any software components, libraries, or solvers used in the experiments. The checklist points to the Appendix for training details, but those are not in the main text. |
| Experiment Setup | Yes | We set the size of the input point cloud to N = 2048. The number and the size of point patches are M = 512 and K = 32, respectively. In the case of Vector Quantization, there are D = 1024 vectors in the dictionary D. The implementation of the uni-directional transformer is based on GPT [42]. It contains 24 blocks, where each block has an attention layer with 16 heads and 1024 embedding dimension. When sampling, nucleus sampling [22] with top-p (0.85) and top-k (100) are applied to predicted token probabilities. |