Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising
Authors: Zikuan Li, Qiaoyun Wu, Jialin Zhang, Kaijun Zhang, Jun Wang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the performance of 3D point cloud denoising using two benchmark datasets, PU-Net (Yu et al. 2018) and PC-Net (Rakotosaona et al. 2020). Ablation Study: We first perform ablation experiments on PU-Net to establish the final architecture of NI-SGCN. All ablation study are performed on the PU-Net dataset, using point clouds with 50K points and 2% Gaussian noise. Table 1: Ablation study on the NI-SGCN architecture designs on the PU-Net dateset. Table 2: Ablation study on the pooling schemes on the PU-Net dataset. Table 3: Ablation study on the usage of different spiking neurons in the network. Table 4: Ablation study on the time latency on the PU-Net dataset. Comparison with State-of-the-art Methods: We compare our method against state-of-the-art deep learning-based denoisers, including PCN (Rakotosaona et al. 2020), Score Denoise (Luo and Hu 2021), DMRDenoise (Luo and Hu 2020), and Pointfilter (Zhang et al. 2020). Table 5: Comparison among competitive denoising algorithms. |
| Researcher Affiliation | Academia | Zikuan Li1, Qiaoyun Wu3*, Jialin Zhang2, Kaijun Zhang2, Jun Wang1,2 1School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics 2College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics 3School of Artificial Intelligence, Anhui University EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods using equations and figures, such as Fig. 1 and Fig. 2 illustrating neuronal dynamics and graph convolution, but it does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Code https://github.com/Miraclelzk/NI-SGCN |
| Open Datasets | Yes | We evaluate the performance of 3D point cloud denoising using two benchmark datasets, PU-Net (Yu et al. 2018) and PC-Net (Rakotosaona et al. 2020). |
| Dataset Splits | Yes | PU-Net contains 60 distinct shapes, which are divided into 40 for training and 20 for testing. The PC-Net dataset, used exclusively for generalization testing, comprises 10 unique point cloud shapes and their corresponding meshes. ... The model is trained on the PU-Net training set, and its denoising performance is evaluated on both the PU-Net test set and the PC-Net dataset. |
| Hardware Specification | Yes | All experiments are implemented on Intel an i9-13900HX CPU and an NVIDIA RTX 4090 GPU (24GB memory, CUDA 11.8) |
| Software Dependencies | No | All experiments are implemented on Intel an i9-13900HX CPU and an NVIDIA RTX 4090 GPU (24GB memory, CUDA 11.8), using Py Torch and Spiking Jelly (Fang et al. 2023) for implementation. The paper mentions CUDA 11.8 with a version, but Py Torch and Spiking Jelly, which are key software components, are mentioned without specific version numbers. |
| Experiment Setup | Yes | NI-SGCN is trained with the Adam optimizer using a learning rate of 1 10 4, and the network is trained with a batch size of 32. For all our SNN models, the default time delay T is set to 4, the membrane potential threshold is set to 1, and the standard deviation of injected noise is 0.2. |