DA-3Det: Monocular 3D Object Detection via Feature Domain Adaptation

August 2020

tl;dr: Use Domain Adaptation to bridge the gap between pseudo-lidar and real lidar.

Overall impression

DA-3Det uses a Siamese network and takes in real lidar and pseudo-lidar data. The difference between the features are penalized. This way DA-3Det learns a general feature based on pseudo-lidar.

Similar ideas to bridge the gap between real and pseudo-lidar has been witnessed in RefinedMPL, which proposes a way to downsample the dense lidar points to mimic the sparsity of point cloud.

Key ideas

Technical details