Semantic3D, S3DIS
收藏资源简介:
本研究利用Semantic3D和S3DIS两个知名LiDAR数据集构建了3D语义分割中分布外检测的基准。Semantic3D作为分布内数据集,因其包含密集的RGB颜色信息而被选用。S3DIS则作为分布外数据集,因其包含室内对象与Semantic3D的室外场景形成对比。这两个数据集通过深度集合和Flipout版本的RandLA-Net生成的最大Softmax概率和熵分数进行评估,旨在提高自动驾驶等安全关键应用中对分布外输入的检测能力。
This study constructs a benchmark for out-of-distribution (OOD) detection in 3D semantic segmentation using two well-known LiDAR datasets: Semantic3D and S3DIS. Semantic3D is selected as the in-distribution (ID) dataset due to its dense RGB color information. S3DIS serves as the out-of-distribution dataset, as its indoor object scenes contrast with the outdoor scenarios of Semantic3D. Both datasets are evaluated using the maximum Softmax probability and entropy scores generated by Deep Ensembles and the Flipout variant of RandLA-Net, aiming to enhance the detection capability of out-of-distribution inputs for safety-critical applications such as autonomous driving.




