STF_dense_fog (SeeingThroughFog)

我们在具有挑战性的恶劣天气条件下引入了一个对象检测数据集，涵盖了真实驾驶场景中的 12000 个样本和雾室内受控天气条件下的 1500 个样本。该数据集包括雾、雪和雨等不同的天气条件，是在北欧超过 10,000 公里的驾驶中获得的。右侧显示了沿途城市的驾驶路线。总共有 100k 个对象用准确的 2D 和 3D 边界框进行标记。该数据集的主要贡献是： - 我们为涵盖信号增强、域适应、对象检测或多模态传感器融合的广泛算法提供了试验场，专注于学习传感器之间的鲁棒冗余，特别是如果它们在不同的天气条件下不对称地失败。 - 创建数据集的最初目的是展示方法，这些方法学习传感器之间的稳健冗余，并在由于恶劣天气影响引起的不对称传感器故障的情况下启用原始数据传感器融合。 - 在我们的案例中，我们脱离了提案级别的融合，并应用了由测量熵驱动的自适应融合，从而在未知的不利天气影响的情况下也能进行检测。该方法优于其他参考融合方法，甚至低于单图像方法。 - 请查看我们的论文以获取更多信息。

We introduce an object detection dataset under challenging adverse weather conditions, which includes 12,000 samples from real-world driving scenarios and 1,500 samples from indoor controlled foggy weather conditions. The dataset covers diverse weather conditions such as fog, snow and rain, and was collected from over 10,000 kilometers of driving in Northern Europe. The driving routes across the surveyed cities are shown on the right. A total of 100k objects are annotated with precise 2D and 3D bounding boxes.

The primary contributions of this dataset are listed below:
- We provide a testbed for a broad spectrum of algorithms including signal enhancement, domain adaptation, object detection and multimodal sensor fusion, with a focus on learning robust redundancy across sensors, particularly when they fail asymmetrically under varying weather conditions.
- The original intent of developing this dataset is to showcase methods that learn robust redundancy between sensors and enable raw-data sensor fusion in cases of asymmetric sensor failures induced by adverse weather effects.
- In our study, we move beyond proposal-level fusion and apply adaptive fusion driven by measurement entropy, allowing detection even under unknown adverse weather conditions. The proposed method outperforms other reference fusion approaches and even outperforms single-image baseline methods.
- Please refer to our paper for additional information.

搜集汇总

背景与挑战

背景概述

STF_dense_fog是一个专注于恶劣天气条件下对象检测的数据集，包含真实驾驶和室内受控环境的样本，总计约13500个样本和100k个标记对象，覆盖雾、雪和雨等多种天气。它旨在为传感器融合、域适应和对象检测算法提供试验场，特别强调学习传感器间的鲁棒冗余以应对不对称故障，由乌尔姆大学于2020年发布。