Gaussian Splatting Lucas-Kanade
Authors: Liuyue Xie, Joel Julin, Koichiro Niinuma, Laszlo A. Jeni
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method excels in reconstructing highly dynamic scenes with minimal camera movement, as demonstrated through experiments on both synthetic and real-world scenes. 1 ... In this section, we first provide the implementation details of the proposed warp field regulation and then validate our proposed method on five dynamic scene datasets captured with different levels of camera movements. Our method outperforms baseline approaches in both static and dynamic camera settings, achieving state-of-the-art results in quantitative and qualitative evaluations. ... Table 1: Quantitative evaluation of novel view synthesis on the Dynamic Scenes dataset. ... Table 2: Quantitative results on the Dy Check dataset. |
| Researcher Affiliation | Collaboration | Liuyue Xie1, Joel Julin1, Koichiro Niinuma2, L aszl o A. Jeni1 1Carnegie Mellon University, 2Fujitsu Research of America EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and textual explanations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | 1Project page: https://gs-lk.github.io |
| Open Datasets | Yes | We evaluate our method using real-world monocular datasets with varying camera motion. These include Dy Check dataset Gao et al. (2022), captured with handheld monocular cameras; Dynamic Scene Yoon et al. (2020), captured by a stationary multi-view 8 camera rig with significant scene motion; Plenoptic Video Li et al. (2021c), captured using a static rig with 21 Go Pro cameras Go Pro (2023); Hypernerf Park et al. (2021b), which captures objects with moving topologies from a moving camera, suitable for quasi-static reconstruction; and sequences from DAVIS 2017 Perazzi et al. (2016) dataset containing near-static monocular videos. |
| Dataset Splits | No | The paper states that it evaluates novel-view synthesis capabilities on various datasets and reports standard metrics, but it does not specify the exact training, validation, or test splits used for these datasets (e.g., percentages, sample counts, or explicit splitting methodologies). |
| Hardware Specification | Yes | The framework is optimized with Adam Kingma & Ba (2015) as with 3DGS Kerbl et al. (2023) on a NVIDIA A100 Tensor Core Nvidia Corporation (2024). |
| Software Dependencies | No | We implement the warp field using Py Torch Paszke et al. (2019), leveraging its Autodiff library for gradient and Jacobian computations. ... We use the Torchdiffeq Chen (2018) library for numerical integration. ... We use Marigold Ke et al. (2024) and RAFT Teed & Deng (2020) for depth and flow maps. The paper mentions software tools and libraries but does not provide specific version numbers for them. |
| Experiment Setup | Yes | The learning rate of the warp field network is empirically set to decay from 8e-4 to 1.6e-6, and Adam s β range is set to (0.9, 0.999). ... For scenes with rich point clouds to start with, we employ 3000 iterations for initialization to refine the canonical structures. Some highly dynamic scenes contain only a few points or no points in the dynamic regions... we extend the warm-up period to 5000-10, 000 to recover as many Gaussians as possible at the dynamic regions. |