ScanNet++
收藏资源简介:
ScanNet++是一个大规模的高保真室内场景数据集,包含460个高分辨率的3D重建场景,每个场景都有密集的语义和实例标注,以及相应的高质量DSLR图像和iPhone RGB-D序列。数据集通过高端激光扫描器以亚毫米级分辨率捕捉每个场景,同时使用3300万像素的DSLR相机和iPhone进行颜色捕捉。此外,场景重建还通过开放词汇进行语义标注,明确标注了语义模糊的场景,以实现全面的语义理解。ScanNet++支持新颖视图合成和3D语义场景理解的新基准,使评估能够针对以前不可用的精确真实世界地面实况进行。
ScanNet++ is a large-scale high-fidelity indoor scene dataset that comprises 460 high-resolution 3D reconstructed scenes. Each scene features dense semantic and instance annotations, paired with corresponding high-quality DSLR images and iPhone RGB-D sequences. The dataset captures every scene using high-end laser scanners at sub-millimeter resolution, and employs 33-megapixel DSLR cameras and iPhones for color data acquisition. Furthermore, scene reconstructions are annotated with open-vocabulary semantics, and semantically ambiguous scenes are explicitly labeled to facilitate comprehensive semantic understanding. ScanNet++ supports novel benchmarks for novel view synthesis and 3D semantic scene understanding, enabling evaluations against previously inaccessible precise real-world ground truth.

- 1ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes慕尼黑工业大学 · 2023年



