Cross-Camera View-Overlap Recognition

Data accompanying the paper titled Cross-Camera View-Overlap Recognition, published in the proceedings of the European Conference on Computer Vision Workshop and presented used for the evaluation of the framework presented in the publication. The dataset consists of image sequence pairs from four scenarios: two scenarios that were collected with both hand-held and chest-mounted cameras – gate and backyard of four sequences each – and two publicly available datasets – TUM-RGB-D SLAM and courtyard from CoSLAM – for a total of ∼28,000 frames (∼25 minutes). The data consisting of images, annotations, and scripts to process existing public sequences. Image sequences are provided for the collected scenarios gate and backyard. We sub-sampled backyard from 30 to 10 fps for annotation purposes. Image sequences for the scenario office can be found at TUM RGB-D SLAM (fr1_desk, fr1_desk2, fr1_room). Scripts to process these sequences as used in the work are provided. The courtyard scenario consists of four sequences. We sub-sampled courtyard from 50 to 25 fps for annotation purposes. Original sequences are available at CoSLAM project website. For all scenarios, we provide i) the annotation of angular distances, Euclidean distances, and overlap ratio of each view pair across camera sequences; ii) the annotation of the calibration (intrinsic) parameters; and iii) the annotation of the camera poses over time for each camera sequence, as automatically reconstructed with the structure-from-motion pipeline, COLMAP, or exploiting the depth data for the office scenario. Camera poses are saved as .txt file for each sequence using the KITTI format. The pose of each frame is represented as a 3x4 matrix (12 parameters) that is converted into a vector by horizontally concatenating the rows of the matrix: [r11 r12 r13 tx r21 r22 r23 ty => [r11 r12 r13 tx r21 r22 r23 ty r31 r32 r33 tz] r31 r32 r33 tz] Values of the parameters are saved in 6 digit floating point numbers as exponential notation. Along with the dataset, we also provide the global features computed by using DeepBit [code] and NetVLAD [code] for each image of all camera sequences. If you use the data, please cite: A. Xompero and A. Cavallaro, Cross-camera view-overlap recognition, International Workshop on Distributed Smart Cameras (IWDSC), European Conference on Computer Vision Workshops, 24 October 2022. ArXiv: https://arxiv.org/abs/2208.11661 Webpage: http://www.eecs.qmul.ac.uk/~ax300/xview/

本数据集配套于发表于欧洲计算机视觉会议（European Conference on Computer Vision, ECCV）研讨会的论文《Cross-Camera View-Overlap Recognition》（跨相机视场重叠识别），用于评估该论文提出的研究框架。
本数据集包含四类场景下的图像序列对：两类由手持与胸戴相机采集的场景——大门（gate）与后院（backyard）场景，各含4个序列；以及两个公开数据源对应的场景——TUM-RGB-D SLAM与CoSLAM中的庭院（courtyard）场景，总计约28000帧图像（时长约25分钟）。数据包含图像文件、标注文件与用于处理现有公开序列的脚本。
针对采集得到的大门与后院场景，我们提供了对应的图像序列。为满足标注需求，我们将后院场景的帧率从30fps降采样至10fps。办公场景（office）的图像序列可从TUM RGB-D SLAM获取（包含fr1_desk、fr1_desk2、fr1_room三个序列），本工作中用于处理这些序列的脚本已随数据集提供。庭院场景包含4个序列，为满足标注需求，我们将其帧率从50fps降采样至25fps，原始序列可从CoSLAM项目官网获取。
针对所有场景，我们提供三类标注内容：
1. 各相机序列间每一对视图的角距离、欧氏距离与视场重叠率标注；
2. 相机内参（intrinsic parameters）标注；
3. 各相机序列随时间变化的相机位姿标注：这些位姿通过运动恢复结构（Structure-from-Motion, SfM）管线COLMAP自动重建得到，或针对办公场景利用深度数据计算得到。相机位姿以KITTI格式保存为每个序列对应的.txt文件，每帧位姿表示为3×4矩阵（含12个参数），将矩阵的各行水平拼接为向量，形式为$[r_{11}, r_{12}, r_{13}, t_x, r_{21}, r_{22}, r_{23}, t_y, r_{31}, r_{32}, r_{33}, t_z]$，参数值以6位浮点数的科学计数法格式存储。
此外，本数据集还提供了所有相机序列的每张图像通过DeepBit与NetVLAD计算得到的全局特征。
若使用本数据集，请引用以下文献：A. Xompero与A. Cavallaro, 《Cross-camera view-overlap recognition》, International Workshop on Distributed Smart Cameras (IWDSC), European Conference on Computer Vision Workshops, 2022年10月24日。ArXiv预印本链接：https://arxiv.org/abs/2208.11661，项目网页：http://www.eecs.qmul.ac.uk/~ax300/xview/

搜集汇总

背景与挑战

背景概述

该数据集用于跨相机视图重叠识别研究，包含来自四个场景的约28,000帧图像序列，提供详细的角距离、欧氏距离、重叠比等注释，以及相机姿态和全局特征，支持多场景评估和计算机视觉应用。