January 2020
tl;dr: Dataset with radar data from proprietary high resolution radar design.
Active learning scheme based on uncertainty sampling using estimated scores as approximation.
- Radar+camera sees more clearly than lidar+camera, for far away objects and for pedestrians. --> However even with radar, the recall is only ~0.5. Too low for real-world application.
- Cross sensor calibration has two steps: camera lidar 2D-3D with checkerboard, and radar lidar 3D-3D relative pose estimation.
- Annotation has "invisible" objects as well associated via temporal reference, but invisible in camera and lidar.
- Dataset
- Estimation of height in this dataset