3D articulated object retrieval using a graph-based representation

Abstract

In this paper, a retrieval methodology for 3D articulated objects is presented that relies upon a graphbased object representation. The methodology is composed of a mesh segmentation stage which creates the Attributed Relation Graph (ARG) of the object along with a graph matching algorithm which matches two ARGs. The graph matching algorithm is based on the Earth Movers Distance (EMD) similarity measure calculated with a new ground distance assignment. The superior performance of the proposed retrieval methodology against state of the art approaches is shown by extensive experimentation that comprise the application of various geometric descriptors representing the components of the 3D objects that become the node attributes of the ARGs as well as alternative mesh segmentation approaches for the extraction of the object parts. The performance evaluation is addressed in both qualitative and quantitative terms.

An overview of partial 3D object retrieval methodologies

“…Very often in the experimental comparisons, a partial 3D object retrieval method is shown to outperform a generic retrieval method, such as the methods of Vranic [73], Papadakis et al [60] and Johnson and Hebert [42]. However, it should be kept in mind that some state-of-the-art partial 3D object retrieval methods aim to address special object classes, as is the case with the method of Agathos et al [2], which focuses on the retrieval of articulated objects, and the methods of Lavoué [46] and Bronstein et al [10], which focus on the retrieval of non-rigid objects.…”

“…Agathos et al [2] proposed a graph-based representation method that decomposes objects using the mesh segmentation method previously introduced by some of the authors [3]. Geodesic extrema of an object are considered as salient points identified by means of the protrusion function, which depends on geodesic distances of all pairs of points in a neighborhood and reaches its local maxima at the tips of mesh protrusions.…”

Multiscale shape context and re-ranking for deformable shape retrieval

“…Results on McGill dataset: In Table 1, we compare the different evaluation results of our method as well as five state-of-the-art methods, including the Covariance Method (CM) [4], Graph-based method [8], PCA-based VLAT (PVLAT) method [10], Hybrid 2D/3D method [35] and Hybrid BoW method [3]. The evaluation results tell us that SA is better than SG and HA for the multiscale descriptor (e.g.…”

“…Smeets et al [7,1] used the spectra of the geodesic distance matrix as shape feature (called SD-GDM). To perform partial shape matching and recognition, many of the researchers turned to the local shape descriptors, like the segmentation based method [8,9], the depth image based method (BF-DSIFT-E [1], PCA-based VLAT [10]). Local descriptors based on diffusion geometry also presented promising properties (e.g.…”

Covariance Descriptors for 3D Shape Matching and Retrieval

“…Table 2 summarizes the retrieval performance of the proposed method on the McGill dataset. We compare our results to various state-of-the-art methods: the Hybrid BoW [20], the PCA-based VLAT approach [35], the graphbased approach of Agathos et al [1], and the hybrid 2D/3D approach of Papadakis et al [27]. Although the proposed method does not consider structural information of shapes, it achieves the best performance on NN and DCG.…”