In autonomous driving, stereo vision-based depth estimation technology can help to estimate the distance of obstacles accurately, which is crucial for correctly planning the path of the vehicle. Recent work has formulated the stereo depth estimation problem into a deep learning model with convolutional neural networks. However, these methods need a lot of post-processing and do not have strong adaptive capabilities to ill-posed regions or new scenes. In addition, due to the difficulty of the labelling the ground truth depth for real circumstance, training data for the system is limited. To overcome the above problems, the authors came up with self-improving pyramid stereo network, which can not only get a direct regression disparity without complicated post-processing but also be robust in ill-posed area. Moreover, by online learning, the proposed model can not only address the data limitation problem but also save the time spent on training and hardware resources in practice. At the same time, the proposed model has a self-improving ability to new scenes, which can quickly adjust the model according to the test data in time and improve the accuracy of prediction. Experiments on Scene Flow and KITTI data set demonstrate the effectiveness of the proposed network.

中文翻译：

---

基于自我完善的金字塔立体网络的智能交通系统深度估算

在自动驾驶中，基于立体视觉的深度估计技术可帮助准确估计障碍物的距离，这对于正确规划车辆的路径至关重要。最近的工作已经将立体深度估计问题公式化为具有卷积神经网络的深度学习模型。但是，这些方法需要大量的后处理，并且对不良区域或新场景没有强大的自适应能力。另外，由于难以为实际情况标记地面真实深度，因此该系统的训练数据受到限制。为了克服上述问题，作者提出了一种自我完善的金字塔立体网络，该网络不仅可以在不进行复杂的后处理的情况下获得直接的回归视差，而且在不适定的区域具有较强的鲁棒性。而且，通过在线学习 提出的模型不仅可以解决数据限制问题，而且可以节省实践中的培训和硬件资源。同时，提出的模型对新场景具有自我完善的能力，可以根据测试数据及时对模型进行快速调整，提高了预测的准确性。在“场景流”和KITTI数据集上进行的实验证明了该网络的有效性。