Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting

Authors: Haiqian Han, Jianing Li, Henglu Wei, Xiangyang Ji

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

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate the effectiveness of our Event-3DGS, we conduct experiments on the Deep Voxels synthetic dataset [38] and the real-world Event-Camera dataset[29]. For the synthetic dataset, we use seven sequences with continuous 180-degree image rotations on a gray background as the ground truth for reconstruction.
Researcher Affiliation Academia Hanqian Han Jianing Li Henglu Wei Xiangyang Ji Tsinghua University Corresponding author: xyji@tsinghua.edu.cn
Pseudocode No The paper describes its method through text, mathematical equations, and a diagram (Figure 1), but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available in https://github.com/lanpokn/Event-3DGS.
Open Datasets Yes To evaluate the effectiveness of our Event-3DGS, we conduct experiments on the Deep Voxels synthetic dataset [38] and the real-world Event-Camera dataset[29].
Dataset Splits No For longer sequences, we typically utilize the initial 100 images for training and evaluate performance on separate data not employed during reconstruction. The paper mentions training on initial images and evaluating on 'separate data' but does not specify clear train/validation/test splits, percentages, or how the separate data is partitioned.
Hardware Specification Yes All experiments are conducted on an AMD Ryzen Threadripper 3970X 32-Core CPU and an NVIDIA GeForce RTX 3080 Ti GPU.
Software Dependencies No The paper mentions using 'E2VID [34]' but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions or specific library versions).
Experiment Setup Yes We set τ to 0.05 for the high-pass filter-based photovoltage contrast estimation module. In the loss function, we set α to 0.9. For synthetic experiments with low noise, β is set to 0, while for real data with higher noise, β is set to 0.5.