PointNet++ — 3D point clouds bounding box detection and tracking (PointNet, PointNet++, LaserNet, Point Pillars and Complex Yolo) — Series 5 (Part 3)
Welcome to this 6 parts series where we discuss five pioneering research papers for you to get started with 3D object detection. In this article, we will discuss Point Net++ by Charles R. Qi. et. al. which was a pioneering work following up on his previous work on PointNet.
PointNet was great but …
There were some issues which are addressed by PointNet++. PointNet fails to capture local structures in the point cloud scenes and hence we have this better version of PointNet. PointNet++ is a hierarchal network that applies PointNet recursively on a nested partitioning of the input point cloud to capture fine-grained patterns in the data. And why just local features, PointNet++ scales the input and combines features from different scales.
Architecture
As you can see, PointNet++ is composed of set-abstraction levels which are further composed of three layers, i.e. Sampling, Grouping, and a PointNet layer. A set abstraction level takes an N × (d + C) matrix as input that is from N points with d-dim coordinates and C-dim point features. It outputs an N⁰ × (d + C⁰ ) matrix of N⁰ subsampled points with…
