Perception of 3D structure and natural scene statistics: The Southampton-York Natural Scenes (SYNS) dataset.
Perception of 3D structure and natural scene statistics: The Southampton-York Natural Scenes (SYNS) dataset.
复制标题
3D 结构和自然场景统计的感知：南安普敦-约克自然场景 (SYNS) 数据集。
DOI：
10.1167/15.12.726
复制
发表时间：
2015
影响因子：
1.8
通讯作者：
Adams W
中科院分区：
文献类型：
--
作者：
Adams W
我们对人类视觉与其运行环境之间的关系感兴趣。为此，南安普顿大学（英国）和约克大学（加拿大）合作构建了南安普顿-约克自然场景（SYNS）公共数据集。为了代表人类经历的多样化环境，我们从英国汉普郡的 19 个室外和 6 个室内场景类别中采样了场景。英国土地利用数据集确定的户外类别包括农田、沿海沙丘、林地、工业区、湿地、住宅区、农场和果园。室内类别包括住宅、剧院、咖啡馆和办公室。每个场景由三种类型的共同配准数据表示：(i) 地面实况 3D 结构：来自激光测距仪 (LiDAR) 的 360 x 135 深度图，(ii) SpheroCam 捕获的高动态范围图像 (360 x 180) 和 (iii) 18 个立体图像对 (35 x 24)，平铺 360 水平全景图，由定制的高分辨率立体图像捕获装备，相机间距与人类平均瞳距相匹配。分析 LiDAR 数据以确定自然场景中倾斜和倾斜的表面姿态分布。计算以每个 LiDAR 点为中心的补丁的表面法线，并通过交叉验证确定最佳补丁大小。总体而言，倾斜和倾斜的关节分布主要由地平面决定。对于地平线以上的高度，其他规律也很明显，包括基本倾斜轴（垂直表面）处的概率密度升高，以及正面平行处的峰值，正如投影几何所预测的那样。我们将这些自然场景统计数据与人类对表面姿态的感知联系起来，并找到了一般对应关系，人类的倾斜感知偏向于地平面，而倾斜感知偏向于额平行。这些结果表明，人类对表面姿态的感知部分取决于我们视觉环境的生态统计数据。
We are interested in the relationship between human vision and the environment in which it operates. To this end, the University of Southampton (UK) and York University (Canada) have collaborated to build the Southampton-York Natural Scenes (SYNS) public dataset. To represent the diverse environments that humans experience, we sampled scenes from 19 outdoor and 6 indoor scene categories across Hampshire, UK. Outdoor categories, identified by the UK Land Use dataset, include cropland, coastal dunes, woodlands, industrial estates, wetlands, residential areas, farms and orchards. Indoor categories include residential, theatres, cafes and offices. Each scene is represented by three types of co-registered data:(i) Ground truth 3D structure: 360 x 135 depth maps from a laser rangefinder (LiDAR),(ii) High dynamic range images (360 x 180) captured by a SpheroCam and (iii) 18 Stereo image pairs (35 x 24), tiling a 360 horizontal panorama, captured by a custom-built high-resolution stereo rig, with camera separation matched to average human interpupillary distance. LiDAR data were analysed to determine the distribution of surface attitude over slant and tilt in natural scenes. Surface normals were computed for patches centred on each LiDAR point, with the optimal patch size determined by cross-validation. Overall, the joint distribution over slant and tilt is dominated by the ground plane. For elevations above the horizon, other regularities are also apparent, including elevated probability density at the cardinal tilt axes (vertical surfaces), and a peak at fronto-parallel, as predicted by the geometry of projection. We relate these natural scene statistics to human perception of surface attitude and find a general correspondence, with human tilt perception biased toward the ground plane and slant perception biased toward fronto-parallel. These results suggest that human perception of surface attitude is governed in part by the ecological statistics of our visual environment.