【摘要】 提出了从任意视角图像中检测视觉目标的新框架,并通过多视角分类器球面(Multi-View Detector Sphere,MVDS)模型对不同视角分类器之间的关系进行建模,以描述多个视角在识别视觉目标过程中的视角关联。首先,对不同视角的特定目标进行建模,对每个视角训练一个分类器;其次,通过将视角球面三角化,将视角球面均匀划分成若干三角面片,面片顶点所代表的视角之间的关系用以刻画不同视角分类器之间的关系。最后,对于来自未训练视角的目标,可通过融合球面上相邻视角分类器的输出给出其正确的检测结果。在多个公共数据集上的实验结果表明了该算法的有效性和准确性。更多还原

【Abstract】 We proposed a novel framework for detecting generic objects from arbitrary viewpoints described by varied object appearances.Our key insight is to exploit the multi-view detection patterns established by a number of detectors from different viewpoints and their relationships through the Multi-View Detector Sphere(MVDS),reflecting the underlying intrinsic structure for detecting multi-view objects.We first modeled the annotated objects from different viewpoints,and then triangulated the sphere into a number of uniformly distributed meshes to represent the explicit correspondences across view detectors.As a result,multi-view objects from untrained viewpoints can be detected by combining the outputs of the adjacent view detectors on the sphere.Our experiments on several public datasets give promising results for the experimental object classes.更多还原