PNeRFLoc: Visual Localization with Point-Based Neural Radiance Fields

Authors: Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, Zhaopeng Cui

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that PNe RFLoc performs the best on the synthetic dataset when the 3D Ne RF model can be well learned, and significantly outperforms all the Ne RF-boosted localization methods with on-par SOTA performance on the real-world benchmark localization datasets.Experiments We first compare our method with various representative and SOTA learning approaches (Sarlin et al. 2021; Moreau et al. 2022a,b; Brachmann and Rother 2021) on both synthetic datasets and real-world datasets. Then, we offer insights into PNe RFLoc through additional ablation experiments.
Researcher Affiliation Academia 1State Key Lab of CAD & CG, Zhejiang University 2Simon Fraser University bmzhao@zju.edu.cn, mluweiyang@outlook.com, maomao6006@gmail.com, zhpcui@gmail.com, bao@cad.zju.cn
Pseudocode No The paper describes its methods in prose and equations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The code and supplementary material are available on the project webpage: https://zju3dv.github.io/PNe RFLoc/.
Open Datasets Yes Cambridge Landmarks (Kendall, Grimes, and Cipolla 2015) contains five outdoor scenes, with 200 to 2000 images captured at different times for each scene.7Scenes (Shotton et al. 2013) contains seven indoor scenes, captured by a Kinect RGB-D sensor.Replica (Straub et al. 2019) contains eight synthetic indoor scenes, commonly used for SLAM evaluation.
Dataset Splits No The paper states for the Replica dataset: 'We follow i MAP (Sucar et al. 2021), using its produced sequences as training images, with an image size of 1200*680 pixels, and randomly generate 50-120 query images.' However, it does not provide specific percentages or counts for training, validation, and test splits across all datasets, nor does it specify the methodology for these splits (e.g., random seed, stratified).
Hardware Specification Yes All our experiments are evaluated on a single NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No The paper mentions software like 'R2D2' and 'Detectron2' but does not specify their version numbers or other software dependencies with specific versions.
Experiment Setup Yes In the structure-based localization stage, the score threshold St is set to 0.7, and the number of RANSAC iterations is set to 20k. During the rendering-based localization stage, we use the Adam optimizer with a learning rate of 0.001.