3D 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.

近年来，3D形状识别引起了很多关注。尽管在基于视图的3D特征描述方面取得了惊人的进步，但是与基于点云的方法相比，以前的基于多视图的方法仍承受着计算效率的负担。为了克服这一局限性，我们提出了一种基于参数化视图学习机制PVLNet的新型轻量级基于多视图的网络，与以前的多视图相比，仅需1/10 FLOP即可实现最新性能基于视图的方法。在参数化视图学习机制的指导下，视图直接作为PVLNet的参数构建，可以通过梯度下降自动优化。当从视图参数生成深度图像时，使用简化的可微分深度图生成器来确保梯度传播。然后，通过全局最大池聚合CNN提取的多视图特征。我们在ModelNet40和ScanObjectNN上的实验结果证明了该方法的优越性能。网络注意力的可视化进一步解释了我们的PVLNet的效率。