ScaleMaster Dataset
收藏资源简介:
ScaleMaster是由韩国大邱庆北科学技术院团队构建的首个专注于评估单目视觉SLAM系统尺度一致性的基准数据集。该数据集包含25条序列,涵盖多层建筑结构(如图书馆、停车场)、长轨迹(最长884米)、重复纹理场景等挑战性环境,数据通过定制采集设备(iPhone 14 Pro+LiDAR)同步获取视觉帧与高精度点云。其创新性在于系统性地揭示了现有基准未覆盖的会话内尺度漂移和会话间尺度模糊问题,通过引入Chamfer距离等三维地图质量指标,为复杂室内场景下的SLAM可靠性研究提供了关键评估工具。
ScaleMaster is the first benchmark dataset dedicated to evaluating the scale consistency of monocular visual SLAM systems, developed by the research team from Daegu Gyeongbuk Institute of Science and Technology (DGIST). This dataset includes 25 sequences spanning challenging indoor environments such as multi-story buildings (e.g., libraries, parking garages), long trajectories with a maximum length of 884 meters, and scenes with repetitive textures. Visual frames and high-precision point clouds are synchronously acquired using a custom data collection platform equipped with an iPhone 14 Pro and a LiDAR sensor. The core innovation of ScaleMaster is that it systematically reveals the intra-session scale drift and inter-session scale ambiguity issues that are not covered by existing benchmark datasets. By introducing 3D map quality evaluation metrics such as Chamfer Distance, this dataset provides a critical assessment tool for research on SLAM reliability in complex indoor scenes.
- 1Have We Mastered Scale in Deep Monocular Visual SLAM? The ScaleMaster Dataset and Benchmark韩国大邱庆北科学技术院·机器人及机电工程系 · 2026年



