Table 7 PointPillars approach
From: Object detection for automotive radar point clouds – a comparison
Model part | Implementation details |
|---|---|
Architecture | Max points N=35 and max pillars P=8000 (edge length 0.5 m). Loss weights adjusted to βobj=2.5,βloc=0.5,βsiz=2,βang=2 (axis-aligned βang=0), and βcls=0.5. Learning rate 3·10−4, 10 anchors: [(6.1 m, 18.0 m), (2.4 m, 15.3 m), (2.8 m, 7.6 m), (1.5 m, 4.4 m), (0.6 m, 1.5 m)]. Anchors used in original form and rotated by 90∘. Pos./neg. IOU thr. 0.5 / 0.2 |
PointPillars++ | Backbone replacement by Darknet-53 |
Class Proposal | Top 200 NMS boxes @ IOU 0.5, proposal equal to points within predicted boxes |