CarFusion

从宽基线、未校准和非同步的相机中观察到的多个动态刚性物体（例如车辆）的快速准确的 3D 重建具有挑战性。一方面，特征跟踪在每个视图中都能很好地工作，但由于视野重叠有限或由于遮挡，很难在多个摄像机之间进行对应。另一方面，深度学习的进步导致强大的检测器可以跨不同的视点工作，但对于基于三角剖分的重建仍然不够精确。在这项工作中，我们开发了一个框架来融合单视图特征轨迹和多视图检测到的部分位置，以显着改善移动车辆的检测、定位和重建，即使在存在强遮挡的情况下也是如此。我们通过重建 3 分钟窗口内通过的 40 多辆车辆，在繁忙的交通路口展示了我们的框架。Fast and accurate 3D reconstruction of multiple dynamic rigid objects (e.g., vehicles) observed from wide-baseline, uncalibrated, and unsynchronized cameras remains highly challenging. On one hand, feature tracking performs well within individual views, yet establishing correspondences across multiple cameras is difficult due to limited field-of-view overlap or occlusions. On the other hand, advances in deep learning have yielded powerful detectors that operate across diverse viewpoints, yet they remain insufficiently accurate for triangulation-based 3D reconstruction. In this work, we develop a framework that fuses single-view feature trajectories and partially estimated positions derived from multi-view detections to substantially improve the detection, localization, and reconstruction of moving vehicles, even under severe occlusions. We demonstrate our framework at a busy traffic intersection by reconstructing over 40 vehicles passing through within a 3-minute window.

搜集汇总

背景与挑战

背景概述

CarFusion是一个发布于2018年的数据集，专注于通过融合单视图特征轨迹和多视图部分检测，实现动态车辆在复杂交通场景下的快速3D重建。该数据集由卡内基梅隆大学发布，旨在提升车辆在遮挡环境中的检测、定位和重建精度。