SAFT: Sensitivity-Aware Filtering and Transmission for Adaptive 3D Point Cloud Communication over Wireless Channels
Authors: Huda Adam Sirag Mekki, Hui Yuan, Mohanad M. G. Hassan, Zejia Chen, Guanghui Zhang
Organizations: School of Control Science and Engineering, Shandong University, Ji’nan, 250061, China · Key Laboratory of Machine Intelligence and System Control, Ministry of Education, Ji’nan, 250061, China · School of Computer Science and Technology, Shandong University, Qingdao, China
Reliable transmission of 3D point clouds over wireless channels is challenging due to time-varying signal-to-noise ratio (SNR) and limited bandwidth. This paper introduces sensitivity-aware filtering and transmission (SAFT), a learned transmission framework that integrates a Point-BERT-inspired encoder, a sensitivity-guided token filtering (STF) unit, a quantization block, and an SNR-aware decoder for adaptive reconstruction. Specifically, the STF module assigns token-wise importance scores based on the reconstruction sensitivity of each token under channel perturbation. We further employ a training-only symbol-usage penalty to stabilize the discrete representation, without affecting the transmitted payload. Experiments on ShapeNet, ModelNet40, and 8iVFB show that SAFT improves geometric fidelity (D1/D2 PSNR) compared with a separate source--channel coding pipeline (G-PCC combined with LDPC and QAM) and existing learned baselines, with the largest gains observed in low-SNR regimes, highlighting improved robustness under limited bandwidth.
Figures & tables
Fig. 1: Overview of the proposed SAFT.
Fig. 2: Detailed architecture of the transformer-based point-token encoder.
Fig. 3: Sensitivity scoring MLP. The module estimates token-wise sensitivity scores S∈[0,1] from Z .
Fig. 4: SNR embedding module.
Fig. 5: Detailed architecture of the refinement-based decoder for 3D point cloud reconstruction.
Fig. 6: Performance comparison on the ShapeNet dataset across SNR levels. (a) D1 PSNR; (b) D2 PSNR.
Method
SEPT
G-PCC (octree)+LDPC
V-PCC+LDPC
PCST
SAFT
AWGN
Rayleigh
AWGN
Rayleigh
AWGN
Rayleigh
AWGN
Rayleigh
AWGN
Rayleigh
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
0 dB
10 dB
Soldier
Metrics
D1 ↑
29.34
30.37
27.83
28.43
–
70.75
–
–
–
71.84
–
67.81
N/A
71.42
N/A
70.72
69.82
71.78
64.87
69.81
D2 ↑
30.54
31.91
28.86
29.59
–
75.55
–
–
–
74.99
–
69.76
N/A
74.82
N/A
74.06
74.58
76.50
69.94
74.43
Avg ↑
29.94
31.14
28.35
29.01
–
73.15
–
–
–
73.42
–
68.79
N/A
73.12
N/A
72.39
72.20
74.14
67.41
72.12
TABLE I: D1 and D2 PSNR on 8iVFB under AWGN and Rayleigh at SNRs of 0 and 10 dB.
Fig. 7: Performance comparison on the ModelNet40 dataset across SNR levels. (a) D1 PSNR; (b) D2 PSNR.
Fig. 8: Robust reconstruction on Longdress under AWGN and Rayleigh channels. (a) D1 (AWGN); (b) D2 (AWGN); (c) D1 (Rayleigh); (d) D2 (Rayleigh).
Fig. 9: Visual comparison of the Andrew under an AWGN channel with SNR = 10 dB.
Fig. 10: Input-density robustness at SNR=−10 dB: N=2048 vs. N=4096 . Colors show normalized pointwise error.
Fig. 11: D1 PSNR across CBR at different SNR levels.
Fig. 12: Ablation study of the SAFT architecture under varying SNRs. (a) D1 PSNR; (b) D2 PSNR.