GenS: Generalizable Neural Surface Reconstruction from Multi-View Images
Authors: Rui Peng, Xiaodong Gu, Luyang Tang, Shihe Shen, Fanqi Yu, Ronggang Wang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on popular benchmarks show that our model can generalize well to new scenes and outperform existing state-of-the-art methods even those employing ground-truth depth supervision. To demonstrate the quantitative and qualitative effectiveness of Gen S, we conduct extensive experiments on DTU [12] and Blended MVS [59] datasets. We conduct ablation studies on DTU dataset to understand how the components of our model contribute to the overall performance. |
| Researcher Affiliation | Collaboration | Rui Peng1,2 Xiaodong Gu3 Luyang Tang1 Shihe Shen1 Fanqi Yu1 Ronggang Wang ,1,2 1School of Electronic and Computer Engineering, Peking University 2Peng Cheng Laboratory 3Alibaba Group |
| Pseudocode | No | The paper describes the proposed methods in detail within the text and uses diagrams to illustrate concepts (e.g., Figure 2, Figure 3, Figure 5, Figure 6), but it does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code will be available at https://github.com/prstrive/Gen S. |
| Open Datasets | Yes | We conduct experiments on both DTU [12] and Blended MVS [59] datasets as previous methods [49, 60, 23]. Our generalization model is trained on DTU dataset, which is an indoor MVS dataset with 124 different scenes scaned from 49 or 64 views with fixed camera trajectories. |
| Dataset Splits | No | The paper mentions a 'training set' and 'test scenes' defined as in prior work, but it does not explicitly provide details about a separate 'validation' split (e.g., percentages, sample counts, or specific predefined validation sets) used for hyperparameter tuning or early stopping during their training process. |
| Hardware Specification | Yes | We train the joint loss for 16 epochs on two A100 GPUs. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer [17]' and an 'FPN network [19]' but does not provide specific version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used for implementation. |
| Experiment Setup | Yes | We use Adam optimizer [17] with the base learning rate of 1e-3 for feature network and 5e-4 for other MLPs. We train the joint loss for 16 epochs on two A100 GPUs. We increase the value of α from 0 to 1 and in the first 2 epochs. ... with β set to 1. We build the generalized multi-scale volume with 5 scales, whose resolution increase from 24 to 28. |