MonoSample: Synthetic 3D Data Augmentation Method in Monocular 3D Object Detection

Abstract

In the context of autonomous driving, it is both critical and challenging to locate 3D objects by using a calibrated RGB image. Current methods typically utilize heteroscedastic aleatoric uncertainty loss to regress the depth of objects, thereby reducing the impact of noisy input while also ensuring the reliability of depth predictions. However, experimentation reveals that uncertainty loss can also lead to serious overfitting issue and performance degradation. To address this issue, we propose MonoSample, an augmentation method that collects samples from the dataset and places them randomly during training. MonoSample takes into account the occlusion relationships and applies strict restrictions to ensure the verisimilitude of the enhanced scenes. Furthermore, MonoSample avoids the complex conversion process between 2D and 3D, thereby enabling the extraction of a large number of samples and efficient operation. Experiments on different models have verified its effectiveness. Leveraging MonoSample in DID-M3D, our model achieves state-of-the-art (SOTA) performance on the KITTI 3D object detection benchmark.