SECOND

SECOND: Sparsely Embedded Convolutional Detection

sensors 2018（机器人顶会） 优化VoxelNet

---

摘要

基于LiDAR或基于RGB-D的目标检测被用于从自动驾驶到机器人视觉的众多应用中。 在处理点云LiDAR数据时，基于Voxel的3D卷积网络被用于提取信息。 但是推理速度慢和方向估计性能低等问题仍然存在。 因此，我们研究了一种针对此类网络的改进的稀疏卷积方法，该方法显着提高了训练和推理的速度。 我们还介绍了一种新的角度损失回归形式，以改善方向估计性能；提出一种新的数据增强方法，可以提高收敛速度和性能。 网络可以在保持快速推理的同时，在KITTI 3D对象检测基准上产生SOTA结果。

问题

试图解决现有3d卷积网络推理速度慢，方向估计精度差的问题。

创新

• 使用改进的空间稀疏卷积网络(Spatially sparse convolutional networks)，从而大大提高了训练和推理的速度。
• 新颖的角度损失回归方法，提高方向回归精度。
• 新颖的数据增强方法，提高了收敛速度和性能。

网络结构

Point Cloud Grouping + Voxelwise Feature Extractor（VFE）(详见VoxelNet)+ Sparse Convolutional Middle Extractor + Region Proposal Network(RPN)

（1）sparse convolution algorithm：

（2）Rule Generation Algorithm：

（3）Sparse Convolutional Middle Extractor

（4）RPN

Loss Function

• Sine-Error Loss for Angle Regression
  

(1) it solves the adversarial example problem between orientations of 0 and p.

(2) it naturally models the IoU against the angle offset function.

• Focal Loss for Classification
  

处理前景背景类别不平衡的问题

• Total Training Loss
  

where L_cls is the classification loss, L_reg-other is the regression loss for location and dimension, L_reg-θ is our novel angle loss, and L_dir is the direction classification loss. β₁ = 1.0, β₂ = 2.0, and β₃ = 0.2 are constant coefficients of our loss formula. We use a relatively small value of β₃ to avoid cases in which our network would struggle to recognize the directions of objects.

数据集

• KITTI数据集扩充：

  First, we generated a database containing the labels of all ground truths and their associated point cloud data (points inside the 3D bounding boxes of the ground truths) from the training dataset. Then, during training, we randomly selected several ground truths from this database and introduced them into the current training point cloud via concatenation. Using this approach, we could greatly increase the number of ground truths per point cloud and simulate objects existing in different environments. To avoid physically impossible outcomes, we performed a collision test after sampling the ground truths and removed any sampled objects that collided with other objects.

实验结果

有无数据扩充的区别

链接

• 项目地址