WildGaussians: 3D Gaussian Splatting In the Wild
Authors: Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc Pollefeys, Torsten Sattler
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our Wild Gaussians approach on two challenging datasets. The Ne RF On-the-go dataset [31] contains multiple casually captured indoor and outdoor sequences, with varying ratios of occlusions (from 5% to 30%). |
| Researcher Affiliation | Academia | 1 Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague 2 Faculty of Electrical Engineering, Czech Technical University in Prague 3 Department of Computer Science, ETH Zurich 4 Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague |
| Pseudocode | No | The paper describes algorithms but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code, model checkpoints, and video comparisons are available at: https://wild-gaussians.github.io/ |
| Open Datasets | Yes | We evaluate our Wild Gaussians approach on two challenging datasets. The Ne RF On-the-go dataset [31] contains multiple casually captured indoor and outdoor sequences, with varying ratios of occlusions (from 5% to 30%). [...] The Photo Tourism dataset [35] consists of multiple 3D scenes of well-known monuments. [...] The Ne RF On-the-go dataset is licensed under the Apache 2.0 license (https://raw.githubusercontent.com/cvg/nerf-on-the-go/refs/heads/master/LICENSE). The pictures in the Photo Tourism dataset were sourced from various creators who made the images available under permissive licenses (https://creativecommons.org/share-your-work/cclicenses/). |
| Dataset Splits | No | The paper describes training steps and evaluation protocols but does not provide specific percentages or counts for training, validation, and test dataset splits for reproduction. |
| Hardware Specification | Yes | All our experiments were conducted on a single NVIDIA RTX 4090 GPU. [...] Methods were trained and evaluated on NVIDIA A100, while the rest used NVIDIA GTX 4090. |
| Software Dependencies | No | The paper mentions software components like 3DGS renderer, Mip-Splatting, Gaussian Opacity Fields, DINO v2, and Adam optimizer, but it does not specify their version numbers. |
| Experiment Setup | Yes | We optimize each representation for 30k training steps. For our choice of learning rates, we mostly follow 3DGS [14]. We differ in the following learning rates: appearance MLP lr. of 0.0005, uncertainty lr. of 0.001. Gaussian embedding lr. of 0.005. image embedding lr. of 0.001. For the position learning rate, we exponentially decay from 1.6 10 4 to 1.6 10 6. Furthermore, we set the densification threshold at 0.0002 and density from 500-th iteration to 15 000th iteration every 100 steps. Furthermore, we reset the opacity every 3 000 steps. We do not optimize the uncertainty predictor for 500 after each opacity reset, and we do not apply the masking for the first 2000 training steps. |