Spiking NeRF: Representing the Real-World Geometry by a Discontinuous Representation

Authors: Zhanfeng Liao, Yan Liu, Qian Zheng, Gang Pan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments, Experimental Settings, Datasets. We evaluate our methods in 8 scenes from the Blender dataset (Mildenhall et al. 2021) and 6 scenes from Dex-Ne RF (Ichnowski et al. 2021). Metrics. Following most 3D reconstruction methods, we measure the reconstruction quality with the Chamfer distances, Overall Performance, Ablation Study.
Researcher Affiliation Academia Zhanfeng Liao, Yan Liu, Qian Zheng*, Gang Pan* Zhejiang University {zhanfengliao, qianzheng, rliuyan2, gpan}@zju.edu.cn
Pseudocode No The paper describes methods using equations and prose, but does not include a clearly labeled "Pseudocode" or "Algorithm" block.
Open Source Code Yes The source code and the supplementary material are available at https://github.com/liaozhanfeng/Spiking-Ne RF.
Open Datasets Yes We evaluate our methods in 8 scenes from the Blender dataset (Mildenhall et al. 2021) and 6 scenes from Dex-Ne RF (Ichnowski et al. 2021). Following most 3D reconstruction methods (Wang et al. 2021; Wang, Skorokhodov, and Wonka 2022, 2023; Oechsle, Peng, and Geiger 2021), we also evaluate our methods in 15 scenes from the DTU dataset (Jensen et al. 2014).
Dataset Splits No The paper mentions the datasets used (Blender, Dex-Ne RF, DTU) but does not provide specific train/validation/test split percentages, sample counts, or explicit references to predefined splits with details. It only cites the datasets.
Hardware Specification Yes We sample 1024 rays per batch and train our model for 400k iterations for 10 hours on a single NVIDIA RTX3090 GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) required to reproduce the experiments.
Experiment Setup Yes We sample 1024 rays per batch and train our model for 400k iterations for 10 hours on a single NVIDIA RTX3090 GPU. Our network architecture and initialization scheme are similar to those of Ne RF (Mildenhall et al. 2021)... We set the initial value of r and k to 100 and 1... we set λ1 to 0.15 at first and continue to increase λ1 as the training process progresses. To preserve true high-frequency geometric information, we initially set the λ2 to 0.0001 and gradually decrease the proportion of Lg in L as the training progresses by decreasing λ2.