Sim-to-Real Domain Adaptation for Lane Detection and Classification in Autonomous Driving

The antitumor activity of a novel GCN2 inhibitor in head and neck squamous cell carcinoma cell lines

“…Comparison with Sim-to-Real adaptation method: The framework using HD map can be regarded as similar to the simulator in that it generates diverse augmented data. Therefore, we compared the performance on TUSimple with existing Sim-to-Real domain adaptation models [13,42,12] in Table IV. Their photo-realistic source domain was made by CARLA [43] which is a gaming engine to generate synthetic datasets.…”

“…Their photo-realistic source domain was made by CARLA [43] which is a gaming engine to generate synthetic datasets. Note that, UNIT [13], MUNIT [42], and ADA [12] models included TUSimple's unlabeled image as they are domain adaptation works. However, our result outperforms the existing Sim-to-Real DA performance without taking advantage of any information from TUSimple.…”

“…Their main goal is to bridge the domain gap between photorealistic synthetic images and real-world data. ADA [12] utilizes adversarial discriminative and generative methods for adapting to the target domain, and UNIT [13] utilizes cycle-GAN-based methods.…”

Self-supervised few-shot learning for real-time traffic sign classification

“…Hu et al [15] present methods of Unsupervised Domain Adaptation (UDA) for identifying and categorizing lanes in self-driving vehicles, utilizing artificial data from virtual settings. Rectifying the mode collapse challenge in adversarial learning methods for unsupervised domain adaptation (UDA) was proposed by Chen et al [16].…”

自动驾驶开源数据体系：现状与未来

“…Labeling Platform [261] SA 丰富的自动驾驶仿真系统被开发出来，如 CARLA [30] 、SUMMIT [256] 、SVL [254] 、PGDrive [255] 、 AirSim [256] 等。这些仿真系统通常包含自定义景观、路网、天气等场景信息，配置不同类型的传感 部分工作尝试将自然语言引入自动驾驶领域中，如 NuInstruct [272] 将基于驾驶视频的问答文本作为 数据标签对外公开，旨在鼓励业界探索语言在自动驾驶中的作用，GAIA-1 [273] 将生成模型与语言结 [280] 提出，伴随 着神经网络、高性能计算、车载传感器等技术的持续演进，在近两年得到迅猛发展。新一代的端到 端自动驾驶技术以 UniAD [281] 、MP3 [282] 、ST-P3 [283] 为代表，采用深度学习框架主干网络整合和 优化自动驾驶全流程的关键任务，从而在感知、预测和规划等方面达到更高的准确性和效率。可以 看出，第一与第二代自动驾驶数据集以感知任务为主，辅以少部分预测和规划任务，不能直接用于 端到端自动驾驶模型的训练与评测。CARLA [30] 提供的仿真环境使得端到端闭环评测成为可能，但…”