ObjectGAN：自动驾驶评估数据集构建

自动驾驶评估数据集的构建在很大程度上取决于能否覆盖复杂多样的交通场景。但是通过人工获取数据工作量巨大，且难以覆盖所有场景，在这种情况下，虚拟引擎可以提供极大的便利。利用UE4（Unreal Engine 4）虚拟引擎构建常见驾驶场景，并结合设计的后处理系统和自动标注算法高效获取、标注复杂场景图像。针对虚拟引擎生成图像逼真度不足这一问题，搭建ObjectGAN域适应模型，基于虚拟数据重建逼真图像，该模型针对目标数据，引入特征一致性监督，无须另外标注信息便可有效缩小与真实数据间域差异。创建了一个新颖的复杂场景虚拟自动驾驶数据集，其中包含多种天候、光照、驾驶场景数据。通过该数据集验证ObjectGAN模型可以有效缩小虚拟数据与真实数据间域差异，经过域适应处理后的数据可以在复杂场景中对主流检测器进行有效的性能评估。

ObjectGAN:Autonomous Driving Evaluation Dataset Construction

The construction of the autonomous driving evaluation dataset depends to a large extent on the ability to reproduce complex traffic scenarios.However,the workload of manually obtaining data is huge and it is difficult to cover all scenarios.In this case,the virtual engine can provide great convenience.UE4 (Unreal Engine 4) virtual engine is used to reconstruct common driving scenes,and post-processing system and automatic labeling algorithm are designed to obtain and label complex scene images efficiently.In order to solve the problem of insufficient image fidelity generated by the virtual engine,an ObjectGAN domain adaptation model is built,and the realistic images is reconstructed based on the virtual data.This model introduces feature consistency supervision for the target data and can reduce the domain shift with the real data without additional labeling information effectively.A novel complex scene virtual autonomous driving dataset is created,which contains a variety of weather,lighting,and driving scene data.This dataset verifies that the ObjectGAN model can effectively reduce the domain difference between virtual data and real data,and the data after domain adaptation processing can evaluate the performance of mainstream detectors in complex scenes effectively.