Search

Abstract In this paper, we address the problem of estimating a 3D human pose from a single image, which is important but difficult to solve due to reasons, such as self-occlusions, wild appearance changes, and inherent ambiguities of 3D estimation from a 2D cue. These difficulties make the problem ill-posed, which have become requiring increasingly complex estimators to enhance the performance. On the other hand, most existing methods try to handle this problem based on a single complex estimator, which might not be good solutions for 3D human pose estimation. In this paper, to resolve this issue, we propose a multiple-partial-hypothesis-based framework for the problem of estimating 3D human pose from a single image, which can be fine-tuned in an end-to-end fashion. We first select several joint groups from a human joint model using the proposed sampling scheme, and estimate the 3D pose of each joint group separately based on deep neural networks. After that, the estimated poses are aggregated to obtain the final 3D pose using the proposed robust optimization formula. The overall procedure can be fine-tuned in an end-to-end fashion, resulting in better estimation performance. In the experiments, the proposed framework shows the state-of-the-art performances on popular benchmark data sets, namely Human3.6M and HumanEva, which demonstrate the effectiveness of the proposed framework. Highlights • A multiple-partial-hypothesis-based scheme for the single-image-based 3D human pose estimation is proposed. • A novel domain conversion layer which transforms a heatmap to corresponding 2D point coordinates is proposed. • The proposed method can be end-to-end fine-tuned to improve the overall performance. • The proposed framework provides the state-of-the-art performance for popular benchmark data sets. [ABSTRACT FROM AUTHOR]

We present a method for three-dimensional (3D) localization and tracking of multiple targets by using images from multiple cameras with overlapping views. Most recent methods make an assumption on a flat ground plane and determine the 3D grounding location of each target based on this assumption. In contrast, we aim to find 3D locations of multiple head trajectories, regardless of their standing or sitting status without the flat ground plane assumption. For this purpose, we suggest a unified optimization formulation, which solves two coupled problems simultaneously: the spatio-temporal data association problem and the 3D trajectory estimation problem. To handle a large solution space, we develop an efficient optimization scheme that alternates the solving procedures between two coupled problems with a reasonable computational load. In the unified optimization formulation, we design a new cost function that describes 3D physical properties of each target. The experiments illustrate that the proposed method outperforms the state-of-the-art methods in 3D localization and tracking performance. [ABSTRACT FROM AUTHOR]

3D human pose estimation from a single image is an important problem in computer vision with a number of applications, including action recognition and scene understanding. However, it is still challenging due to its ill-posedness and complex non-rigid shape variations of a human body. In this paper, we use the Procrustean normal distribution mixture model as a 3D shape prior and propose a model transformation method for adjusting limb lengths of the 3D shape prior model, by which the proposed method can be applied to a novel test image. Inaccuracies of 2D part detections are handled by selecting from a diverse set of 2D pose candidates considering both the 2D part model and 3D shape model. Experimental results show that the proposed method performs favorably compared with existing methods, despite inaccuracies of 2D part detections and 3D shape ambiguities. [ABSTRACT FROM AUTHOR]

Highlights: [•] We propose EM-GPA algorithm to handle shapes with missing data. [•] EM GPA combines GPA and the expectation-maximization (EM) algorithm. [•] 2D shapes can be considered as 3D shapes with missing depth information. [•] EM-GPA finds scales, rotations and 3D shapes along with their mean and covariance. [•] The method accurately estimates the 3D mean and covariance matrix for 3D shape modeling. [ABSTRACT FROM AUTHOR]

Abstract: We present a new tracking method with improved efficiency and accuracy based on the subspace representation and particle filter. The subspace representation has been successfully adopted in tracking, e.g., the Eigen-tracking algorithm, and it has shown considerable robustness for tracking an object with changing appearance. Particle filters are widely used for a wide range of tracking problems since they can efficiently handle non-Gaussian and nonlinearity. Their combination has shown superior performance in terms of accuracy and robustness, but it suffers from the heavy computational load. Our tracking algorithm requires a significantly small number of particles while maintaining robustness and accuracy. We propose two methods in our tracking algorithm: first, we analyze object motion in a coarse-to-fine way and use hierarchical strategy to estimate it, in which the Kalman filter estimates global linear motion and the particle filter handles the local nonlinear motion, second, we give a more physically meaningful proposal distribution of the particle filter with consideration of the nature of motion. Experiments demonstrate the effectiveness of our tracking algorithm in real video sequences in which the target objects undergo rapid and abrupt motion. Furthermore, we provide quantitative comparisons between the existing tracking algorithm and the proposed tracking algorithm. [Copyright &y& Elsevier]

Recent object detectors localize instances and classify candidate regions simultaneously. The number of candidate regions is typically larger than the number of objects and each region is evaluated independently. To assign a single detection bounding box for each object, heuristic algorithms, such as non-maximum suppression (NMS), have been used widely. While simple heuristic algorithms are effective for stand-alone objects, they often fail to detect overlapped objects. In this paper, we address this issue by training a network to distinguish different objects using the relationship between candidate boxes. We propose an instance-aware detection network (IDNet), which can learn to extract features from candidate regions and measure their similarities. Based on pairwise similarities and detection qualities, the IDNet selects a subset of candidate bounding boxes using instance-aware determinantal point process inference (IDPP). Extensive experiments demonstrate that the proposed algorithm achieves significant improvements for detecting overlapped objects compared to existing state-of-the-art detection methods on CrowdHuman, Pascal VOC, and MS COCO datasets. • The greedy NMS, which has been widely used as a post processing method in object detection, has a poor performance in crowd scenes. • To alleviate this problem, we designed a new trainable post processing method using Determinantal Point Processes. • For training, we proposed an instance-aware detection (ID) loss and sparse score (SS) loss to learn features and scores of objects, respectively. • We demonstrate that the proposed algorithm outperforms baseline detectors for detecting overlapped objects. [ABSTRACT FROM AUTHOR]

Nonrigid registration finds transformations to fit a source point cloud/mesh to a target point cloud/mesh. Most nonrigid registration algorithms consist of two steps; finding correspondence and optimization. Among these, finding correspondence plays an important role in registration performance. However, when two point clouds have large displacement, it is hard to know correct correspondences and an algorithm often fails to find correct transformations. In this paper, we propose a novel graph-matching-based correspondence search for nonrigid registration and a corresponding optimization method for finding transformation to complete nonrigid registration. Considering global connectivity as well as local similarity for the correspondence search, the proposed method finds good correspondences according to semantics and consequently finds correct transformations even when the motion is large. Our algorithm is experimentally validated on human body and animal datasets, which verifies that it is capable of finding correct transformations to fit a source to a target. • We propose a graph-matching model for nonrigid registration and its solver. • Input 3D data is divided into small rigid pieces to find correspondences. • We use graph-matching for finding coarse correspondences under large motions. • We propose a many-to-one matching problem to find semantically identical parts robustly. • Experiments show that the proposed method outperforms existing algorithms. [ABSTRACT FROM AUTHOR]