Timing

This page contains the typical timing that we note for relevant pre-processing and GPU inference steps, which are used when deploying this code in a ROS node.

All times are in ms.

Preprocessing

• Sample Dict: 20-22
  • Create BEV: 12-15
  • (optional) Ground plane estimation: <1
  • Anchors:
    • Create: <1 (pre-generated positions and sizes)
    • (optional) Filter: 8-9

GPU Inference

AVOD

• Total: 60
  • RPN: 50
  • Second stage: 7.5
  • Python overhead: 2.5

AVOD-FPN

• Total: 80