Abstract | ||
---|---|---|
AbstractHighlights •We propose parameterized-view-learning mechanism which aims to parameterize the views as parameters in the multi-view networks.•We provide an intuitionistic interpretation of parameterized-view-learning mechanism by analyzing point cloud networks as multi-view networks.•We propose a novel efficient multi-view network based on parameterized-view-learning mechanism, called PVLNet.•Our model can achieve state-of-the-art performance with 1/10 parameters compared with previous multi-view networks.•The experimental results on synthetic and real-world 3D datasets demonstrate the effective performance of the proposed PVLNet. Graphical abstractWe propose parameterized-view-learning (PVL) mechanism to build an efficient and light-weight multi-view based network, named as PVLNet. From the experiments on ScanObjectNN and ModelNet40 benchmark, with 1/10 FLOPS and GPU runtime compared with the state-of-the-art methods, our PVLNet can achieve great performance and superior generalization ability.Display OmittedAbstract3D shape recognition has drawn much attention in recent years. Despite the amazing progress on view-based 3D feature description, previous multi-view based methods suffer from a burden in computation efficiency compared with point cloud based methods. To overcome the limitation, we propose a novel light-weight multi-view based network built on parameterized-view-learning mechanism, PVLNet, which can achieve the state-of-the-art performance with only 1/10 FLOPs compared with previous multi-view based methods. Guided by the parameterized-view-learning mechanism, the views are directly built as parameters of PVLNet which can be automatically optimized by gradient descent. A simplified differentiable depth map generator is used to ensure the gradient propagation when generating depth images from view parameters. Then multi-view features extracted by CNNs are aggregated by global max-pooling. Our experimental results on ModelNet40 and ScanObjectNN demonstrate the superior performance of the proposed method. The visualization of the networks attention further interprets the efficiency of our PVLNet. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1016/j.cag.2021.04.036 | Periodicals |
Keywords | DocType | Volume |
Point cloud, Multi-view, 3D Shape recognition | Journal | 98 |
Issue | ISSN | Citations |
C | 0097-8493 | 0 |
PageRank | References | Authors |
0.34 | 0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Hongbin Xu | 1 | 0 | 1.69 |
Lvequan Wang | 2 | 0 | 0.34 |
Qiuxia Wu | 3 | 9 | 3.20 |
Wenxiong Kang | 4 | 38 | 6.66 |