Sequence Matters: Harnessing Video Models in 3D Super-Resolution

Authors: Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, Eunbyung Park

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results show that the surprisingly simple algorithms can achieve the state-of-the-art results of 3D super-resolution tasks on standard benchmark datasets, such as the Ne RF-synthetic and Mip Ne RF-360 datasets.
Researcher Affiliation Collaboration 1Department of Electrical and Computer Engineering, Sungkyunkwan University 2Department of Artificial Intelligence, Sungkyunkwan University 3Visual Display Division, Samsung Electorics
Pseudocode Yes Algorithm 1: A Simple Greedy Algorithm Input: A set of unordered images, I = {Ij}N j=1 Output: An ordered sequence of images, S
Open Source Code Yes Project Page https://ko-lani.github.io/Sequence-Matters
Open Datasets Yes Datasets We use the Ne RF Synthetic Blender dataset (Mildenhall et al. 2021) and the Mip-Ne RF 360 dataset (Barron et al. 2022).
Dataset Splits No The paper mentions using the Ne RF Synthetic Blender dataset and the Mip-Ne RF 360 dataset and downsampling them, but it does not specify the training/test/validation splits or percentages used for experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions using "the open-source 3D Gaussian Splatting code base" and "PSRT" as a VSR backbone, but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Following the 3DGS protocol, we train both coarse and fine 3DGS models for 30,000 iterations. To create the low-resolution (LR) dataset, we downsample the high-resolution (HR) dataset using bicubic interpolation with a downscale factor of 4.