Investigating Augmentation Techniques for Camera-Radar Object Detection

In autonomous driving, a variety of sensors are used for the perception of the surroundings. Commonly, camera and LiDAR sensors are used due to their availability and details in the sensor data. Radar sensors have been gaining popularity due to the radar’s sensor robustness in harsh weather conditions. Datasets with radar data are essential for training accurate models for radar-based sensor fusion algorithms in autonomous driving. Perception neural network algorithms rely on data for training and evaluation and often suffer from a class imbalance in the data as well. This paper uses augmentation on existing data, nuScenes dataset to address the class imbalance problem. The annotations of minority classes of train split were duplicated to a specific multiple. The performance was evaluated on nuScenes validation dataset, which showed improvements in the performance on most of the minority classes.