DeepICP: An End-to-End Deep Neural Network for 3D Point Cloud Registration 2019
该论文其实是feature-based的匹配方法, 不过都是用DL完成的. 既然这是feature-based的方法,那么理论上就可以处理全局匹配的问题。下面是该论文的主要流程图, 跟feature-based的匹配方法像. 论文没有开源代码. 看到这篇论文,我个人比较好奇的点主要就是输入的点云的order会影响结果吗, 为啥叫DeepICP,跟ICP有啥关系? 我们一起来理解一下这篇论文.
- FeatureExtraction 用的是PointNet++(也是一篇非常有名的工作,代码开源啦)的3D点特征提取层. 主要分三步:
a)The Sampling layer selects a set of points from input points, which defines the centroids of local regions.
b) Grouping layer then constructs local region sets by finding “neighboring” points around the centroids. 这一步可以解决点云order的问题
c)PointNet layer uses a mini-PointNet to encode local region patterns into feature vectors.
2. PointWeighting. Inspired by the attention layer in 3DFeatNet, we design a point weighting layer to learn the saliency of each point in an end-to-end framework. Ideally, points with invariant and distinct features on static objects should be assigned higher weights.
3. DeepFeatureEmbeding. The DFE layer we used is a mini-PointNet structure. 这里主要是结合邻近的region来进一步提取特征.
4. CorrespondingPointGeneration.
our approach utilizes the powerful generalization capability of CNNs in similarity learning, to directly “guess” where the corresponding points are in the target point cloud. 换一句话说,就是“我们用了feature匹配.”
5. PoseEstimation. 论文有监督的方法.
论文对比的传统的匹配方法都是icp的方法,这里我觉得不太公平, 应该对比一下FGR和TEASER这类feature-based的方法。
我不太知道怎么解读DL的论文,因为我不会设计, 只会用别人的框架排列组合,哈哈哈哈。