DualMLP: a two-stream fusion model for 3D point cloud classification

In this paper, we present DualMLP, a novel 3D model that introduces the idea of a two-stream network for existing 3D models to handle the trade-off between the number of points and the computational overhead. Existing works on point clouds use a small subset of points sampled from the entire 3D object as input. Although increasing the number of input points can enhance scene understanding, it also incurs a higher computational cost for existing networks. To tackle this challenge, we propose a novel architecture called DualMLP, which effectively mitigates the linear increase in computational expense as the number of input points grows. While we evaluate this concept on PointMLP and demonstrate its effectiveness, the idea can be applied to other existing models with minimal adjustments. DualMLP consists of two branches: DenseNet and SparseNet. The SparseNet, a relatively larger network, samples a small number of points from the complete point cloud, while the DenseNet, a lightweight network, takes in a larger number of points as input. Extensive experiments on the ScanObjectNN and ModelNet40 datasets demonstrate the effectiveness of the proposed model, achieving a 1.00% and 0.81% improvement over PointMLP for ScanObjectNN and ModelNet40 while being computationally efficient than the original PointMLP. To ensure the reproducibility of our experimental results, the code for this work is publicly available at https://github.com/snehaputul/DualMLP.