Rethinking Mesh Watermark: Towards Highly Robust and Adaptable Deep 3D Mesh Watermarking
Authors: Xingyu Zhu, Guanhui Ye, Xiapu Luo, Xuetao Wei
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate our method remains efficient and effective even if the mesh size is increased. |
| Researcher Affiliation | Academia | Southern University of Science and Technology, Shenzhen 518055, China 2Department of Computer Science and Engineering, Southern University of Science and Technology, China 3Department of Computing, Hong Kong Polytechnic University, Hong Kong |
| Pseudocode | No | The paper describes the architecture and training process in detail but does not include a formal pseudocode block or algorithm. |
| Open Source Code | No | The paper does not provide an explicit statement or link to its own source code for the described methodology. |
| Open Datasets | Yes | Our DE EP3DMA RK is trained on a simplified train set from Model Net40 (Wu et al. 2015) and then tested on the entire test of Model Net40 and other datasets such as Shape Net (Chang et al. 2015), Grasp Net (Fang et al. 2020), Scan Net (Dai et al. 2017) and Hands (Romero, Tzionas, and Black 2022). |
| Dataset Splits | No | The paper specifies training and test set sizes (e.g., "3508 train meshes and 879 test meshes for m2500"), but it does not mention a distinct validation set or its split. |
| Hardware Specification | Yes | Our experiment is conducted on Ubuntu 18.04, with 503GB RAM and five Nvidia RTX 3090. |
| Software Dependencies | Yes | Our experiment is conducted on Ubuntu 18.04, with 503GB RAM and five Nvidia RTX 3090. We generated two train sets m500 and m2500. The number of vertices in m500 and m2500 are Nv = 500 and Nv = 2500, respectively. For m2500, we manually filter out meshes whose Nv is originally less than 2500 and those with low quality after simplifications. We also perform the same process for m500. As a result, we get 3508 train meshes and 879 test meshes for m2500, and 1147 train meshes and 337 test meshes for m500. The original Model Net has 9843 and 2468 meshes for training and testing. We train two replicas of DEEP3DMARK on m500 and m2500, respectively. Both are further tested on the test set of Model Net to evaluate the size adaptability. To evaluate the effectiveness on geometry variations, two replicas of DEEP3DMARK, which is trained on m500 and m2500, are tested on Shape Net, Grasp Net, Scan Net, and Hands. Shape Net has different categories of meshes from Model Net, such as birdhouse, camera, clock, etc. Scannet is a dataset of scanned and reconstructed real-world scenes. Hands contain meshes of human hands. We acquire simplified data through a simplification using CGAL (The CGAL Project 2022), which performs edge-collapse or half-edge-collapse algorithms to reduce the number of triangles by merging vertices. The CGAL Project. 2022. CGAL User and Reference Manual. CGAL Editorial Board, 5.5.1 edition. |
| Experiment Setup | Yes | During training, we set λenc = 2, λdec = 1, λdis = 0.001 under the settings of 8-bit message lengths, and we set µ = 0, σ = 0.001, α [0, π), s [0.1, 1). |