VCR-GauS: View Consistent Depth-Normal Regularizer for Gaussian Surface Reconstruction
Authors: Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, Gim Hee Lee
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Compared with Gaussian-based baselines, experiments show that our approach obtains better reconstruction quality and maintains competitive appearance quality at faster training speed and 100+ FPS rendering. |
| Researcher Affiliation | Academia | 1 School of Computing, National University of Singapore 2 Princeton University hanlin.chen@u.nus.edu gimhee.lee@nus.edu.sg |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Our codes have been released. (NeurIPS Paper Checklist, Q5). The project website linked in the paper (https://hlinchen.github.io/projects/VCR-GauS/) states "Code & Data (Coming soon)". |
| Open Datasets | Yes | For surface reconstruction, we evaluate on Tanks and Temples (TNT) [24]. To further validate the effectiveness of our method, we compare with other methods on Replica [43]. Although we focus on the large-scale reconstruction, we also report our results on DTU [21], which can be seen in the supplementary. Furthermore, we evaluate the rendering results on Mip-Ne RF360 [3]. |
| Dataset Splits | Yes | We use the same train and test data with 2DGS on TNT, Mip-Ne RF360, and DTU datasets. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA 3090/4090/A5000/A6000 GPU. |
| Software Dependencies | Yes | We use PyTorch 2.0.1 and CUDA 11.8. |
| Experiment Setup | Yes | λ1, λ2, and λ3 are set to 1, 0.01, and 0.015, respectively. The densification threshold β is set to 0.002. The hyperparameter γ is set to 0.005. We use pretrained DSINE [1] to predict normal maps for outdoor scenes and pretrained Geo Wizard [13] for indoor scenes. Similar to 3DGS, we stop densification at 15k iterations and optimize all of our model parameters for 30k iterations. |