PORF: POSE RESIDUAL FIELD FOR ACCURATE NEURAL SURFACE RECONSTRUCTION

Authors: Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, Philip Torr

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct a comprehensive evaluation on both DTU Jensen et al. (2014) and Mobile Brick Li et al. (2023a) datasets. We conduct extensive ablation studies on the DTU dataset Jensen et al. (2014).
Researcher Affiliation Academia Jia-Wang Bian Department of Engineering Science University of Oxford Wenjing Bian Department of Engineering Science University of Oxford Victor Adrian Prisacariu Department of Engineering Science University of Oxford Philip Torr Department of Engineering Science University of Oxford
Pseudocode No The paper describes its methods verbally and with mathematical equations, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper states: "Moreover, we integrate our method into the Nerfstudio library, leading to consistently improved performance in diverse challenging scenes." However, this refers to using an existing open-source library and does not explicitly state that the authors' specific implementation of Po RF or their full method's code is open-source or provide a link to it.
Open Datasets Yes We perform our experiments on both the DTU Jensen et al. (2014) and Mobile Brick Li et al. (2023a) datasets.
Dataset Splits No The paper mentions using "all available images" for training and 15 or 18 "test scenes" for evaluation, but it does not explicitly provide details about a separate validation split, its size, or how it was created, which is needed to reproduce the data partitioning for validation.
Hardware Specification Yes The entire framework undergoes training for 50,000 iterations, which takes 2.5 hours in one NVIDIA A40 GPU.
Software Dependencies No The paper mentions using the "Nerfstudio library" and building upon "Neu S" and "Voxurf", but it does not provide specific version numbers for these software components or any other libraries/dependencies required to replicate the experiments.
Experiment Setup Yes To calculate the proposed epipolar geometry loss, we randomly sample 20 image pairs in each iteration and distinguish inliers and outliers by using a threshold of 20 pixels. The entire framework undergoes training for 50,000 iterations... Specifically, we randomly sample 512 rays from a single image in each iteration, with 128 points sampled per ray.