SLNet is a super-lightweight PyTorch framework for 3D point cloud classification and segmentation. The architecture integrates Nonparametric Adaptive Point Embedding (NAPE), Lightweight Geometric Modulation Units (GMU), parameter-free normalization, and compact residual MLP refinement.
The design objective is maximum accuracy-per-parameter efficiency. SLNet achieves strong performance with an extremely low parameter count, minimal memory footprint, and low computational complexity—eliminating the need for attention mechanisms, heavy graph convolutions, or dense residual stacks.
This makes SLNet particularly suitable for edge devices, embedded GPUs, and resource-constrained systems such as NVIDIA Jetson platforms.
The SLNet pipeline featuring the NAPE Front-End, GMU, and four encoder stages.
Nonparametric Adaptive Point Embedding maps raw 3D coordinates using fused Gaussian RBF and Cosine harmonic bases.
Geometric Modulation Unit performs lightweight geometric feature modulation via channel-wise gating.
Measured on NVIDIA Jetson Orin Nano (1024 points, batch size 4).
| Dataset | Model | GFLOPs | Param (M) | Mem (MB) | Time (ms) |
|---|---|---|---|---|---|
| ModelNet40 | SLNet-S | 0.31 | 0.14 | 21.30 | 17.83 |
| ModelNet40 | SLNet-M | 1.22 | 0.55 | 33.00 | 30.11 |
| ScanObjectNN | SLNet-S | 0.26 | 0.12 | 17.87 | 16.36 |
| ScanObjectNN | SLNet-M | 1.02 | 0.48 | 27.37 | 27.43 |
We propose two composite efficiency indices to benchmark model performance against computational constraints:
$$ NetScore = 20 \log_{10} \left( \frac{a^2}{\sqrt{p \cdot m}} \right) $$
$$ NetScore+ = 20 \log_{10} \left( \frac{a^2}{\sqrt{p \cdot m} \cdot (t \cdot r)^{1/4}} \right) $$
Where a = accuracy, p = parameters, m = FLOPs, t = latency, and r = peak memory. SLNet consistently dominates these metrics.
@article{saeid2026slnet,
title={SLNet: A Super-Lightweight Geometry-Adaptive Network for 3D Point Cloud Recognition},
author={Saeid, Mohammad and Salarpour, Amir and MohajerAnsari, Pedram and Pes{\'e}, Mert D},
booktitle={2026 IEEE International Conference on Robotics and Automation (ICRA)},
year={2026}
}