Surface Snapping Optimization Layer for Single Image Object Shape Reconstruction
Authors: Yuan-Ting Hu, Alex Schwing, Raymond A. Yeh
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we evaluate the proposed surface snapping technique on the challenging Pix3D (Sun et al., 2018) and Shape Net (Chang et al., 2015) datasets. On both datasets, we observe that surface snapping, when incorporated into recent baselines, improves the reconstruction; especially in terms of the normal consistency metric. Qualitatively, we observe that the reconstructed geometry is less noisy, i.e., the surface is smoother. This demonstrates the effectiveness of the proposed surface snapping layer at adjusting face normals for monocular shape reconstruction. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign 2Department of Computer Science, Purdue University. |
| Pseudocode | No | The paper describes computational steps and mathematical formulations but does not include any explicitly labeled “Pseudocode” or “Algorithm” blocks. |
| Open Source Code | No | The paper does not contain any explicit statement or link indicating the release of source code for the described methodology. |
| Open Datasets | Yes | In experiments, we evaluate the proposed surface snapping technique on the challenging Pix3D (Sun et al., 2018) and Shape Net (Chang et al., 2015) datasets. |
| Dataset Splits | No | We use the rendered images provided by Choy et al. (2016) and use the training and test splits provided by Wang et al. (2018). The training split consists of 840,189 images and the test split contains 210,051 images." The paper explicitly mentions train and test splits, but does not explicitly mention or quantify a validation split. |
| Hardware Specification | Yes | We compare the following settings (on an Nvidia Tesla V100 GPU): |
| Software Dependencies | No | Leveraging the sparsity structure and solving with torch.solve takes 1.71s and consumes 4437Mi B in GPU memory;" - only generic software names are mentioned without specific version numbers (e.g., 'torch.solve' without PyTorch version). |
| Experiment Setup | No | Following Gkioxari et al. (2019), we train both the model parameters of the refinement modules and the surface snapping strength α in Eq. (2) by minimizing the sum of a Chamfer loss Lcham, the normal distance Lnorm, and an edge regularizer Ledge, i.e., L = Lcham + λ1Lnorm + λ2Ledge." The paper describes the loss function components and mentions training the strength α, but does not provide specific numerical values for hyperparameters such as learning rate, batch size, or the values of λ1 and λ2. |