Probabilistic Implicit Scene Completion
Authors: Dongsu Zhang, Changwoon Choi, Inbum Park, Young Min Kim
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 EXPERIMENTS We test the probabilistic shape completion of c GCA for scenes (Section 4.1) and single object (Section 4.2). Table 1: Quantitative comparison of probabilistic scene completion in Shape Net scene dataset with different levels of completeness. |
| Researcher Affiliation | Academia | Dongsu Zhang, Changwoon Choi, Inbum Park & Young Min Kim Department of Electrical and Computer Engineering, Seoul National University 96lives@snu.ac.kr, changwoon.choi00@gmail.com, {inbum0215, youngmin.kim}@snu.ac.kr |
| Pseudocode | No | The paper outlines the training procedure in a numbered list (Section 3.3, points 1 and 2) but does not present a formal pseudocode block or an algorithm figure. |
| Open Source Code | Yes | Code to run the experiments is available at https://github.com/96lives/gca. |
| Open Datasets | Yes | We evaluate our method on two datasets: Shape Net scene (Peng et al. (2020)) and 3DFront (Fu et al. (2021)) dataset. We use chair, sofa, table classes of Shape Net (Chang et al. (2015)). |
| Dataset Splits | Yes | The number of rooms for train/val/test split are 3750/250/995. Thus, 7044/904/887 rooms are collected as train/val/test rooms. |
| Hardware Specification | No | The paper discusses 'GPU usage' and 'efficient GPU memory usage' (Appendix D, Appendix A.2) but does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for the experiments. |
| Software Dependencies | No | We use Pytorch (Paszke et al. (2019)) and the sparse convolution library Minkowski Engine (Choy et al. (2019)) for all implementations. We train the autoencoder with Adam (Kingma & Ba (2015)) optimizer. The specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | For all the experiments, we use latent code dimension K = 32, trained with regularization parameter β = 0.001. We train the autoencoder with Adam (Kingma & Ba (2015)) optimizer with learning rate 5e-4 and use batch size of 3, 1, and 4 for Shape Net scene (Peng et al. (2020)), 3DFront (Fu et al. (2021)), and Shape Net (Chang et al. (2015)) dataset, respectively. The transition kernel is trained using infusion schedule αt = 0.1 + 0.005t and standard deviation schedule σt = 10 − 1 0.01t. Lastly, we use mode seeking steps T = 5 for all experiments. |