Understanding the cognition ability of a model is a very important task in machine learning systems. Although deep learning algorithms have been able to learn a variety of mapping relationships based on the data, these mappings are not always accurate, leading to uncertainty of the model output. In autonomous driving, the perception uncertainty should be seriously considered, which can significantly impact the subsequent tasks such as vehicle behaviour planning and control. This presentation discusses about 3D object detection approaches in the traffic environment by considering uncertainty in the neural networks. The focus is on the probabilistic modeling of the mapping between the 2D visual representation and the 3D spatial attributes. State-of-the-art performance on public benchmarks is also presented.