Ground-aware Monocular 3D Object Detection for Autonomous Driving (EN)
This is my paper accepcted by RAL 2021. The open-sourced code is in https://github.com/Owen-Liuyuxuan/visualDet3D .
The basic idea is an attempt to mimic how people perceive depth from a single image, and more importantly try to incorperate calibration matrix and ground plane information into the detection model.
Core Operations and Code Placement
- Precomputing statistics for anchors: script github page
- Using the statistics for anchors: head github page
- Ground-Aware Convolution Module: block github page
- Change the "cfg.detector.name" in config to Yolo3D and experiment with DeformConv (which also provide robust and top performance).
Result for the published model:
Benchmark | Easy | Moderate | Hard |
---|---|---|---|
Car Detection | 92.35 | 79.57 | 59.61 |
Car Orientation | 90.87 | 77.47 | 57.99 |
Car 3D Detection | 21.60 | 13.17 | 9.94 |
Car Bird's Eye View | 29.38 | 18.00 | 13.14 |