GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting

Authors: Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, Renjie Wan

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on the Blender, LLFF, and Mip Ne RF-360 datasets to validate the effectiveness of our proposed method, demonstrating state-of-the-art performance on both message decoding accuracy and view synthesis quality.
Researcher Affiliation Collaboration Xiufeng Huang1,2, Ruiqi Li1, Yiu-ming Cheung1, Ka Chun Cheung2, Simon See2, Renjie Wan1 1 Department of Computer Science, Hong Kong Baptist University 2 NVIDIA AI Technology Center
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Project page: https://kevinhuangxf.github.io/Gaussian Marker.
Open Datasets Yes Dataset. We use three benchmark datasets for evaluation: Blender [7] (8 detailed synthetic objects), LLFF [53] (9 real-world scenes), and Mip-Ne RF360 [54] (9 real-world scenes).
Dataset Splits Yes For Blender [7], we directly follow the dataset splitting to use 100 viewpoints for training and 200 views for testing. For LLFF [53], we follow the dataset splitting in Ne RF [7]. In general, 1/8 images in each scene are used for testing and others for training. For Mip-Ne RF360 [54], we use a train/test split suggested by Mip-Ne RF360, taking every 8th photo for testing and others for training.
Hardware Specification Yes The training takes 1000 (Blender, LLFF) or 2000 steps (Mip Ne RF360) and can finish within 20 minutes using a single NVIDIA V100 GPU.
Software Dependencies No The paper mentions using the 'Adam optimizer [24]' but does not provide specific version numbers for any programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or other key software libraries used in the experiments.
Experiment Setup Yes We use the Adam optimizer [24] to optimize the 3D message decoder Dψ and classifier Dξ with default values β1 = 0.9, β2 = 0.999, ϵ = 10 8, and a learning rate 1 10 4 that decays following the exponential scheduler during optimization. We set λ1 = 10.0, λ2 = 1.0 for 2D watermarking loss L2D and λ 1 = 2.0, λ 2 = 1.0 for 3D watermarking loss L3D to adapt the training losses.