3dod_review.md

A Survey on 3D Object Detection Methods for Autonomous Driving Applications

October 2019

tl;dr: Summary of 3DOD methods based on monocular images, lidars and sensor fusion methods of the two.

Overall impression

The review is updated as of 2018. However there have been a lot of progress of mono 3DOD in 2019. I shall write a review of mono 3DOD soon.

Key ideas

• The main drawback of monocular methods is the lack of depth cues, which limits detection and localization accuracy, especially for far and occluded objects.

• Most mono 3DOD methods have shifted towards a learned paradigm for RPN and second stage of of 3D model matching and reprojection to obtain 3D Box.

Technical details

• Mono

  • mono3D
  • 3DVP
  • subCNN
  • deepMANTA
  • deep3DBox
  • 360 panorama

• Lidar

  • projection
    • VeloFCN
    • Complex YOLO
    • Towards Safe (variational dropout)
    • BirdNet (lidar point cloud normalization)
  • Volumetric
    • 3DFCN
    • Vote3Deep
  • point net
    • VoxelNet

• Sensor Fusion

  • MV3D
  • AVOD
  • Frustum PointNet

Notes

• camera and lidar calibration with odometry