3D-BSLS-6D (3D scans of Bins by Structured-Light Scanner for 6D pose estimation)

数据集由手动注释的 Photoneo PhoXi 结构光扫描仪设备的真实捕获和自定义生成器生成的合成样本组成。与现有的 6D 姿态估计数据集相比，一些显着的差异包括：大多数捕获的 bin 是无纹理的，由统一的单色材料制成，所有 bin 都是具有不同比例的长方体形状。与具有复杂几何形状的对象相比，bin 由带有边缘的平面组成，由于遮挡，不能保证在捕获中看到它们。没有提供这些箱的表面模型，只是它们的近似边界框，PhoXi 扫描仪提供高分辨率 3D 几何数据，但没有 RGB 数据，粗糙和嘈杂的灰度强度图像是最接近的等效值，捕获来自不同的设备具有各种内在相机参数。与 RGBD 图像相反，3D 点云隐含地包含这些参数。由于其目前的大小有限，我们建议使用交叉验证而不是显式的训练验证拆分。我们计划在数据集中添加更多样本。注意：应忽略后缀为 _bad、_ish 或 _catastrophic 的注释文件。使用我们当前的工具集无法正确注释它们。

This dataset comprises real-captured samples from the manually annotated Photoneo PhoXi structured-light scanner and synthetic samples generated by a custom generator. Compared to existing 6D pose estimation datasets, several notable distinctions include: Most of the captured bins are textureless, fabricated from uniform monochromatic materials, and all bins are cuboid-shaped with varying scales. Unlike objects with complex geometric structures, bins are composed of flat surfaces with edges, which cannot be guaranteed to be visible in captures due to occlusion. No surface models of these bins are provided, only their approximate bounding boxes. The PhoXi scanner delivers high-resolution 3D geometric data but no RGB data; rough and noisy grayscale intensity images act as the closest available equivalent. Captures originate from different devices with diverse intrinsic camera parameters. In contrast to RGBD images, these intrinsic parameters are implicitly embedded within the 3D point clouds. Owing to its currently limited scale, we recommend employing cross-validation rather than an explicit train-validation split. We intend to add more samples to the dataset. Note: Annotation files with suffixes _bad, _ish, or _catastrophic should be disregarded, as they cannot be properly annotated using our current toolset.

搜集汇总

背景与挑战

背景概述

3D-BSLS-6D是一个用于6D姿态估计的数据集，包含通过结构光扫描仪捕获的真实数据和合成样本，主要针对无纹理、单色材料的长方体形状bin。该数据集提供高分辨率3D点云但缺乏RGB信息，与常见数据集在纹理和几何上存在差异，建议采用交叉验证方法进行使用。