Learning-Deep-Learning

Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving

Mar 2019

tl;dr: Estimate depth map from RGB image (mono/stereo) and use it to lift RGB to point cloud. The point cloud representation uses depth information more efficiently and steps toward bridging the gap between camera based approach and lidar based approach.

Overall impression

This paper opens up a brand new field for camera perception. It points out that the current inefficiency in 3D object detection based on RGB/D image. This idea is highly thought-provoking and we can build upon this idea for new types of hardware setup for autonomous driving. This work seems to be inspired by MLF based on mono images.

The Uber ATG group also publishes several papers (ContFuse, MMF) on this idea, although not as explicit as the pseudo-lidar paper or this one.

Note that DORN’s training data overlaps with object detection’s validation data, and suffers from overfitting. Both pseudo lidar and pseudo lidar e2e suffer from this problem. According to the ForeSeE paper, if the validation data is excluded from the training of depth map, then PL’s performance drops from 18.5 to 5.4 AP_3D at IoU=0.7.

In refined Monocular PseudoLidar, the authors sparsified the dense pseudo-lidar point cloud for faster processing, and gained accuracy as well.

Key ideas

Technical details

Notes