Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Enhancing Close-up Novel View Synthesis via Pseudo-labeling

Authors: Jiatong Xia, Libo Sun, Lingqiao Liu

AAAI 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments demonstrate the efficacy of our approach.
Researcher Affiliation Academia Australian Institute for Machine Learning, The University of Adelaide EMAIL
Pseudocode No The paper describes methods using mathematical equations and prose, but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/Jiatong Xia/Close-up-View Pseudo-Labeling.git
Open Datasets No Furthermore, we have developed a dataset specifically designed to evaluate the generation of close-up views. This dataset addresses the current lack of benchmarks for assessing the performance of existing and future methods in this domain.
Dataset Splits Yes Our dataset comprises diverse scenes, each one is extracted as a subset of frames from a captured video, with 50 to 100 training images and 10 to 20 testing images for each scene, images are captured at a resolution of 960 540.
Hardware Specification Yes The experiments are conducted on NVIDIA 3090 GPUs and the Adam optimizer (Kingma and Ba 2015) is employed to optimize the radiance field.
Software Dependencies No The paper mentions the Adam optimizer but does not specify any software libraries or frameworks with version numbers used for implementation.
Experiment Setup Yes After that, for Ne RF as baseline method, we load the weights of the pre-trained Ne RF model and optimized for 10K iterations per scene with a ray batch of original training samples and generated samples set to 2048 (1024 for each). For 2DGS as baseline method, we optimized the pre-trained 2DGS model for 500 iterations per scene with each iteration combine supervision from both training image and masked pseudo label. For test-time finetuning, we applied our method on both Ne RF and 2DGS. Specifically, we optimized the pre-trained 2DGS model with original training samples and generated samples together for iterations of 5 times for each test view. And for Ne RF, we optimized the pre-trained model for 200 iterations on test poses, with a batch size of 1024 for each samples. We randomly select the value of λ in Eq. 8 from the range of (2, 8) for each iteration. And set the value of ε in Eq. 10 to π 4. ϵ denotes the threshold for determining whether two RGB values can be considered as matched, and we set ϵ to 0.05 in this paper.