Meta-Sim

Meta-Sim是由英伟达和多伦多大学等机构联合开发的数据集，旨在通过学习生成模型来合成与真实世界数据分布相匹配的标签数据集。该数据集通过图形引擎生成图像及其对应的地面实况，利用神经网络参数化数据集生成器，以最小化模拟输出与目标数据之间的分布差异。Meta-Sim特别优化了下游任务的性能，适用于从对象检测到车道估计等多种应用场景，旨在解决机器学习中数据收集和标注的成本和时间瓶颈问题。

Meta-Sim is a dataset co-developed by NVIDIA, the University of Toronto and other institutions. It aims to synthesize labeled datasets that match the real-world data distribution by learning generative models. This dataset generates images and their corresponding ground truths via graphics engines, and uses neural networks to parameterize the dataset generator to minimize the distribution discrepancy between simulated outputs and target data. Meta-Sim specifically optimizes the performance of downstream tasks, and is applicable to a wide range of application scenarios ranging from object detection to lane estimation. It is designed to address the cost and time bottlenecks in data collection and annotation for machine learning.