NI-SGCN: Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising (AAAI‘25)

We design noise-injected spiking graph convolutional networks (NI-SGCN) for 3D point cloud denoising, which is based on an ANN-based denoising architecture, ScoreDenoise. The NI-SGCN consists of two main parts: the spiking feature extraction module and the score estimation module.

Consequently, we propose two implementations of NI-SGCN: the first is a hybrid architecture, NI-HSGCN, witch employs the above spiking feature extraction module and the ANN-based score estimation module; the second is NI-PSGCN, a purely SNN-based structure, which fully leverages the enhanced energy efficiency of SNNs.

Environment

• Python 3.8
• PyTorch 2.1.2
• CUDA and CuDNN (CUDA 11.8 & CuDNN 8.7)
• point_cloud_utils 0.30.4
  • For evaluation only. It loads meshes to compute point-to-mesh distances.
• pytorch3d 0.7.5
• For evaluation only. It computes point-to-mesh distances.
• pytorch-cluster 1.6.3
• We only use fps (farthest point sampling) to merge denoised patches.
• spikingjelly 0.0.0.0.14
• other:
  • tqdm
  • scipy
  • scikit-learn
  • pyyaml
  • easydict
  • tensorboard
  • pandas

Datasets

Download link: https://drive.google.com/drive/folders/1--MvLnP7dsBgBZiu46H0S32Y1eBa_j6P?usp=sharing

Please extract data.zip to data folder.

Train

Training of NI_HSGCN

cd models_NI_HSGCN
bash train_snn_Noisy.sh NI-HSGCN-T4 4

workspace=NI-HSGCN-T4

time step t=4

Please find tunable parameters in the script.

Training of NI_PSGCN

cd models_NI_PSGCN
bash train_snn_Noisy.sh NI-PSGCN-T4 4

Please find tunable parameters in the script.

Denoise-test

# PCNet dataset, 10K Points, 0.01 noise, niters 1, T 4
cd models_NI_HSGCN
bash test_snn.sh NI-HSGCN-T4 ../Exp/NI-HSGCN-T4/checkpoint/ckpt_0.000361_974000.pt PCNet 10000_poisson 0.01 1 4

Please find tunable parameters in the script.

Test sequence

cd models_NI_HSGCN
bash seq_test_IF_T4.sh

Citation

If you find this repo useful, please consider citing:

Acknowledgements

This code largely benefits from following repositories:

• ScoreDenoise
• spikingjelly