KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences

Authors: Keng-Wei Chang, Zi-Ming Wang, Shang-Hong Lai

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we compare our approach with existing joint refinement models: BARF , CF3DGS , Nope-Ne RF , and SC-Ne RF , using the Tanks and Temples and CO3DV2 datasets. We also conduct an ablation study to highlight key components of our method. Moreover, we show that our method outperforms 3DGS , even when it uses camera pose estimates from COLMAP , which is often regarded as ground truth to evaluate the effectiveness of pose estimation.
Researcher Affiliation Academia Keng-Wei Chang, Zi-Ming Wang, Shang-Hong Lai National Tsing Hua University EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes methods using mathematical equations and textual explanations but does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about releasing source code, nor does it provide any links to a code repository.
Open Datasets Yes In this section, we compare our approach with existing joint refinement models: BARF , CF3DGS , Nope-Ne RF , and SC-Ne RF , using the Tanks and Temples and CO3DV2 datasets.
Dataset Splits Yes For the CO3DV2 dataset, we followed the experimental setup in CF3DGS, evaluating the same five selected sequences and presenting the results in Tables 3. This dataset is more challenging due to complex trajectories and blurred images. Our method achieves a 4 d B higher average PSNR than others and significantly reduces the training time cost.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions various software tools and methods like COLMAP, DPT, 3D Gaussian Splatting (3DGS), NeRF, BARF, CF3DGS, Nope-Ne RF, SC-Ne RF, but does not specify any version numbers for these or any other software dependencies.
Experiment Setup No The paper mentions using a photometric loss function balanced by a factor λ, and a keyframe subsampling interval of 5. However, it does not provide specific numerical values for hyperparameters such as λ, learning rates, batch sizes, or detailed schedules for the coarse-to-fine frequency-aware densification filter scale σ.