RT3D: Real-Time 3-D Vehicle Detection in LiDAR Point Cloud for Autonomous Driving
IEEE ROBOTICS AND AUTOMATION LETTERS, 2018
RT3D
本方法是個兩階段模型，總體來說就是使用R-FCN在柵格上進行檢測。本文時間較早，所以方法上感覺比較簡單。
體素化
0.05x0.05的體素，統計高度最大值，最小值和均值。不包括點的用（0,0,0）補充。
The point cloud is projected onto a ground-plane grid with resolution 0.05×0.05 m, where each grid cell records (min(z),ave(z),max(z)) of relevant projected points. Grid cells with no point clouds are assigned a triple of (0, 0, 0).
KITTI的場景一般是70mx80m，這樣看，體素化之後的大小是1400x1600的大小。
檢測網絡
基於ResNet-50的R-FCN。作者將R-FCN在提取ROI之前的卷積，K2C個通道的feature map解釋爲通過車輛不同位置預測車輛的位置、輪廓。另外作者提到要注意的兩點是
Two issues caused by the sparsity of the point cloud need to be addressed: 1) during sliding window on feature maps, there are many anchors which have no data. To reduce computation, we delete these empty anchors; 2) many region proposals contain no vehicles or many region proposals contain simple examples, so online hard example mining (OHEM) [19] is adopted to automatically select hard examples to make training more effective and efficient.
實驗結果
精度不高，但是快。Ablation Study也是在討論一些超參數的設計問題。