Decoupling Appearance Variations with 3D Consistent Features in Gaussian Splatting
Authors: Jiaqi Lin, Zhihao Li, Binxiao Huang, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Xiaofei Wu, Fenglong Song, Wenming Yang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our method on several appearancevariant scenes, and demonstrate that it achieves state-of-the-art rendering quality with minimal training time and memory usage, without compromising rendering speeds. Experiments conducted on both our new dataset and existing datasets demonstrate that DAVIGS achieves state-of-the-art performance in rendering quality, optimization time, video memory, and rendering speeds. We evaluate rendering quality using PSNR, SSIM, and LPIPS, and assess efficiency using optimization time, VRAM usage, and rendering speed. Ablation Studies: To assess the effectiveness of our introduced modules, we conduct ablation experiments on the GLAV dataset. |
| Researcher Affiliation | Collaboration | 1Tsinghua University 2Huawei Noah s Ark Lab 3The University of Hong Kong 4Shenzhen Institute of Advanced Technology |
| Pseudocode | No | The paper describes the method conceptually and with a pipeline diagram (Figure 3), but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about the release of source code, a link to a code repository, or mention of code in supplementary materials. |
| Open Datasets | Yes | Existing datasets for evaluating this task are limited: some contain only global and mild appearance variations, such as the Mill-19 dataset (Turki, Ramanan, and Satyanarayanan 2022), which primarily includes camera ISP variations; others contain other distractors, like the Photo Tourism dataset (Jin et al. 2021), which includes both appearance changes and transient objects, alongside with varying image resolutions and qualities. |
| Dataset Splits | Yes | Following the common practice, we use the left half of the test images during training to optimize the appearance-related parameters and reserve the right half for testing. |
| Hardware Specification | Yes | All the metrics are tested on a single Tesla V100 GPU. |
| Software Dependencies | No | The paper mentions that DAVIGS is implemented based on 3DGS, but it does not specify any software libraries or tools with their version numbers. |
| Experiment Setup | Yes | The MLP f contains two hidden layers with widths of 128 and 64, respectively. The hyperparameter λ1 is set to 0.2, while λ2 follows a cosine annealing function, warming up linearly from 0 to 0.3 in the first 5k iterations and eventually decaying to 0.2. For cell-based query, we use a cell size of 8. For GS-based methods, we optimize the models for 30k iterations, keeping other parameters the same as those in the original papers. |