SUNCG
收藏资源简介:
从单视图深度图观察中生成场景的体积占用和语义标签的完整 3D 体素表示的任务。以前的工作已经分别考虑了深度图的场景完成和语义标记。然而,我们观察到这两个问题是紧密交织在一起的。为了利用这两个任务的耦合性质,我们引入了语义场景完成网络(SSCNet),这是一个端到端的 3D 卷积网络,它以单个深度图像作为输入,同时输出所有体素的占用率和语义标签。相机视锥体。我们的网络使用基于扩张的 3D 上下文模块来有效地扩展感受野并实现 3D 上下文学习。为了训练我们的网络,我们构建了 SUNCG——一个手动创建的具有密集体积注释的合成 3D 场景的大规模数据集。我们的实验表明,联合模型优于单独处理每个任务的方法,并且优于语义场景完成任务的替代方法。
The task of generating a complete 3D voxel representation of a scene with volume occupancy and semantic labels from single-view depth map observations. Prior works have separately considered scene completion and semantic labeling for depth maps. However, we observe that these two tasks are tightly intertwined. To leverage the coupled nature of these two tasks, we introduce Semantic Scene Completion Network (SSCNet), an end-to-end 3D convolutional network that takes a single depth image as input and simultaneously outputs the occupancy and semantic labels for all voxels within the camera frustum. Our network employs a dilated 3D context module to effectively expand the receptive field and enable 3D contextual learning. To train our network, we construct SUNCG, a large-scale dataset of synthetic 3D scenes manually created and annotated with dense volume annotations. Our experiments demonstrate that our joint model outperforms methods that handle each task separately, as well as alternative approaches for the semantic scene completion task.

- SUNCG数据集首次发表,由普林斯顿大学和斯坦福大学的研究团队共同开发,旨在提供一个大规模的室内场景合成数据集,以支持计算机视觉和机器人技术的研究。
- SUNCG数据集首次应用于计算机视觉领域的研究,特别是在场景理解、物体识别和三维重建等方面,展示了其在复杂室内环境中的应用潜力。
- SUNCG数据集被广泛应用于机器人导航和路径规划的研究中,为机器人提供了丰富的室内环境数据,促进了相关技术的发展。
- SUNCG数据集的扩展版本发布,增加了更多的室内场景和物体类别,进一步丰富了数据集的内容,提升了其在多领域研究中的应用价值。
- SUNCG数据集在虚拟现实和增强现实领域的研究中得到应用,为创建更真实的虚拟环境提供了数据支持。
- 1SUNCG: A Large-Scale Scene Understanding and Modeling DatasetPrinceton University · 2017年
- 2Learning to Segment Human by Watching YouTubeUniversity of California, Berkeley · 2018年
- 3Learning to Look Around Objects for Top-View Representations of Outdoor ScenesUniversity of California, Berkeley · 2018年
- 4Learning to Segment Every ThingFacebook AI Research · 2018年
- 5Learning to See in the DarkStanford University · 2018年



