On 3-D scene flow and structure recovery from multiview image sequences

Two novel systems computing dense three-dimensional (3-D) scene flow and structure from multiview image sequences are described in this paper. We do not assume rigidity of the scene motion, thus allowing for nonrigid motion in the scene. The first system, integrated model-based system (IMS), assumes that each small local image region is undergoing 3-D affine motion. Non-linear motion model fitting based on both optical flow constraints and stereo constraints is then carried out on each local region in order to simultaneously estimate 3-D motion correspondences and structure. The second system is based on extended gradient-based system (EGS), a natural extension of two-dimensional (2-D) optical flow computation. In this method, a new hierarchical rule-based stereo matching algorithm is first developed to estimate the initial disparity map. Different available constraints under a multiview camera setup are further investigated and utilized in the proposed motion estimation. We use image segmentation information to adopt and maintain the motion and depth discontinuities. Within the framework for EGS, we present two different formulations for 3-D scene flow and structure computation. One formulation assumes that initial disparity map is accurate, while the other does not. Experimental results on both synthetic and real imagery demonstrate the effectiveness of our 3-D motion and structure recovery schemes. Empirical comparison between IMS and EGS is also reported.     

Improved estimation of defocus blur and spatial shifts in spatial domain: a homotopy-based approach

This paper presents a homotopy-based algorithm for the recovery of depth cues in the spatial domain. The algorithm specifically deals with defocus blur and spatial shifts, that is 2D motion, stereo disparities and/or zooming disparities. These cues are estimated from two images of the same scene acquired by a camera evolving in time and/or space. We show that they can be simultaneously computed by resolving a system of equations using a homotopy method. The proposed algorithm is tested using synthetic and real images. The results confirm that the use of a homotopy method leads to a dense and accurate estimation of depth cues. This approach has been integrated into an application for relief estimation from remotely sensed images.     

A Variational Framework for Structure from Motion in Omnidirectional Image Sequences

We address the problem of depth and ego-motion estimation from omnidirectional images. We propose a correspondence-free structure-from-motion problem for sequences of images mapped on the 2-sphere. A novel graph-based variational framework is first proposed for depth estimation between pairs of images. The estimation is cast as a TV-L1 optimization problem that is solved by a fast graph-based algorithm. The ego-motion is then estimated directly from the depth information without explicit computation of the optical flow. Both problems are finally addressed together in an iterative algorithm that alternates between depth and ego-motion estimation for fast computation of 3D information from motion in image sequences. Experimental results demonstrate the effective performance of the proposed algorithm for 3D reconstruction from synthetic and natural omnidirectional images.     

Realtime Object-aware Monocular Depth Estimation in Onboard Systems

This paper proposes the object depth estimation in real-time, using only a monocular camera in an onboard computer with a low-cost GPU. Our algorithm estimates scene depth from a sparse feature-based visual odometry algorithm and detects/tracks objects’ bounding box by utilizing the existing object detection algorithm in parallel. Both algorithms share their results, i.e., feature, motion, and bounding boxes, to handle static and dynamic objects in the scene. We validate the scene depth accuracy of sparse features with KITTI and its ground-truth depth map made from LiDAR observations quantitatively, and the depth of detected object with the Hyundai driving datasets and satellite maps qualitatively. We compare the depth map of our algorithm with the result of (un-) supervised monocular depth estimation algorithms. The validation shows that our performance is comparable to that of monocular depth estimation algorithms which train depth indirectly (or directly) from stereo image pairs (or depth image), and better than that of algorithms trained with monocular images only, in terms of the error and the accuracy. Also, we confirm that our computational load is much lighter than the learning-based methods, while showing comparable performance.

     

基于形变模型由立体序列图象恢复物体的3D形状

结合立体视觉和形变模型提出了一种新的物体3D形状的恢复方法。采用立体视觉方法导出物体表面的3D坐标;利用光流模型估计物体的3D运动,根据此运动移动形变模型,使其对准物体的表面块;由形变模型将由各幅图象得到的离散的3D点融为一起,得到物体的表面形状。实验结果表明该方法能用于形状复杂的物体恢复。     

Structure from Motion: Beyond the Epipolar Constraint

The classic approach to structure from motion entails a clear separation between motion estimation and structure estimation and between two-dimensional (2D) and three-dimensional (3D) information. For the recovery of the rigid transformation between different views only 2D image measurements are used. To have available enough information, most existing techniques are based on the intermediate computation of optical flow which, however, poses a problem at the locations of depth discontinuities. If we knew where depth discontinuities were, we could (using a multitude of approaches based on smoothness constraints) accurately estimate flow values for image patches corresponding to smooth scene patches; but to know the discontinuities requires solving the structure from motion problem first. This paper introduces a novel approach to structure from motion which addresses the processes of smoothing, 3D motion and structure estimation in a synergistic manner. It provides an algorithm for estimating the transformation between two views obtained by either a calibrated or uncalibrated camera. The results of the estimation are then utilized to perform a reconstruction of the scene from a short sequence of images.The technique is based on constraints on image derivatives which involve the 3D motion and shape of the scene, leading to a geometric and statistical estimation problem. The interaction between 3D motion and shape allows us to estimate the 3D motion while at the same time segmenting the scene. If we use a wrong 3D motion estimate to compute depth, we obtain a distorted version of the depth function. The distortion, however, is such that the worse the motion estimate, the more likely we are to obtain depth estimates that vary locally more than the correct ones. Since local variability of depth is due either to the existence of a discontinuity or to a wrong 3D motion estimate, being able to differentiate between these two cases provides the correct motion, which yields the least varying estimated depth as well as the image locations of scene discontinuities. We analyze the new constraints, show their relationship to the minimization of the epipolar constraint, and present experimental results using real image sequences that indicate the robustness of the method.     

Semi-dense and robust image registration by shift adapted weighted aggregation and variational completion

This work presents a novel approach for both stereo and optical flow that deals with large displacements, depth/motion discontinuities and occlusions. The proposed method comprises two main steps. First, a novel local stereo matching algorithm is presented, whose main novelty relies in the block-matching aggregation step. We adopt an adaptive support weights approach in which the weight distribution favors pixels that share the same displacement with the reference one. State-of-the-art methods make the weight function depend only on image features. On the contrary, the proposed weight function depends additionally on the tested shift, by giving more importance to those pixels in the block-matching with smaller cost, as these are supposed to have the tested displacement. Moreover, the method is embedded into a pyramidal procedure to locally limit the search range, which helps to reduce ambiguities in the matching process and saves computational time. Second, the non-dense local estimation is filtered and interpolated by means of a new variational formulation making use of intermediate scale estimates of the local procedure. This permits to keep the fine details estimated at full resolution while being robust to noise and untextured areas using estimates at coarser scales. The introduced variational formulation as well as the block-matching algorithm are robust to illumination changes. We test our algorithm for both stereo and optical flow public datasets showing competitive results.     

自适应成本量的抗遮挡光场深度估计算法

目的 光场相机通过一次成像同时记录场景的空间信息和角度信息,获取多视角图像和重聚焦图像,在深度估计中具有独特优势。遮挡是光场深度估计中的难点问题之一,现有方法没有考虑遮挡或仅仅考虑单一遮挡情况,对于多遮挡场景点,方法失效。针对遮挡问题,在多视角立体匹配框架下,提出了一种对遮挡鲁棒的光场深度估计算法。方法 首先利用数字重聚焦算法获取重聚焦图像,定义场景的遮挡类型,并构造相关性成本量。然后根据最小成本原则自适应选择最佳成本量,并求解局部深度图。最后利用马尔可夫随机场结合成本量和平滑约束,通过图割算法和加权中值滤波获取全局优化深度图,提升深度估计精度。结果 实验在HCI合成数据集和Stanford Lytro Illum实际场景数据集上展开,分别进行局部深度估计与全局深度估计实验。实验结果表明,相比其他先进方法,本文方法对遮挡场景效果更好,均方误差平均降低约26.8%。结论 本文方法能够有效处理不同遮挡情况,更好地保持深度图边缘信息,深度估计结果更准确,且时效性更好。此外,本文方法适用场景是朗伯平面场景,对于含有高光的非朗伯平面场景存在一定缺陷。     

Moving object detection and segmentation in urban environments from a moving platform

This paper proposes an effective approach to detect and segment moving objects from two time-consecutive stereo frames, which leverages the uncertainties in camera motion estimation and in disparity computation. First, the relative camera motion and its uncertainty are computed by tracking and matching sparse features in four images. Then, the motion likelihood at each pixel is estimated by taking into account the ego-motion uncertainty and disparity in computation procedure. Finally, the motion likelihood, color and depth cues are combined in the graph-cut framework for moving object segmentation. The efficiency of the proposed method is evaluated on the KITTI benchmarking datasets, and our experiments show that the proposed approach is robust against both global (camera motion) and local (optical flow) noise. Moreover, the approach is dense as it applies to all pixels in an image, and even partially occluded moving objects can be detected successfully. Without dedicated tracking strategy, our approach achieves high recall and comparable precision on the KITTI benchmarking sequences.     

Monocular scene flow estimation via variational method

Scene flow provides the 3D motion field of point clouds, which correspond to image pixels. Current algorithms usually need complex stereo calibration before estimating flow, which has strong restrictions on the position of the camera. This paper proposes a monocular camera scene flow estimation algorithm. Firstly, an energy functional is constructed, where three important assumptions are turned into data terms derivation: a brightness constancy assumption, a gradient constancy assumption, and a short time object velocity constancy assumption. Two smooth operators are used as regularization terms. Then, an occluded map computation algorithm is used to ensure estimating scene flow only on un-occluded points. After that, the energy functional is solved with a coarse-to-fine variational equation on Gaussian pyramid, which can prevent the iteration from converging to a local minimum value. The experiment results show that the algorithm can use three sequential frames at least to get scene flow in world coordinate, without optical flow or disparity inputting.     

基于变分光流估计的肺部4D-CT图像超分辨率重建

由于受到扫描时间和照射剂量的限制,肺部4D-CT数据中纵向采样率远小于面内采样率.为了得到更高质量的肺部图像,从医学图像固有的自相似性出发,提出了一种基于局部和全局相结合的变分光流估计的图像序列超分辨率重建技术,用于提高4D-CT图像重建质量.首先,构建了一个用于求解肺部4D-CT不同相位图像之间的光流场的变分光流模型;然后,利用快速交替方向乘子法求解该模型,得到不同相位图像之间的光流场;最后,基于光流场,并利用非局部迭代反投影超分辨率重建算法,实现了高分辨率肺部图像的重建.实验结果表明：与已有算法相比,本方法在增强图像纹理结构的同时更好地保留了图像的轮廓.     

A Robust Approach for Structure from Planar Motion by Stereo Image Sequences

This paper proposes a robust method for recovery of motion and structure from two image sequences taken by stereo cameras undergoing a planar motion. The feature correspondences between images are extracted and refined automatically by the relation of the stereo cameras and the property of the motion. To improve the robustness, an auto-scale random sample consensus (RANSAC) algorithm is adopted in the motion and structure estimation. Unlike other work recovering epipolar geometry, here we use a random sampling algorithm to recover the 2D motion and to exclude the outliers which lie both on and out of the epipolar lines. Further more, the idea of RANSAC is used in structure estimation to exclude the outliers from the image sequence. The contribution of this work is the development of an approach to make structure and motion estimation more robust and efficient so as to be applicable to real applications. With the adoption of the auto-scale technique, the algorithm completely automates the estimation process without any prior information or user’s specification of parameters like thresholds. Indoor and outdoor experiments have been done to verify the performance of the algorithm. The results demonstrated that the proposed algorithm is robust and efficient for applications in planar motions.     

A Stochastic Approach for Blurred Image Restoration and Optical Flow Computation on Field Image Sequence

The blur in target images caused by camera vibration due to robot motion or hand shaking and by object(s) moving in the background scene is different to deal with in the computer vision system.In this paper,the authors study the relation model between motion and blur in the case of object motion existing in video image sequence,and work on a practical computation algorithm for both motion analysis and blut image restoration.Combining the general optical flow and stochastic process,the paper presents and approach by which the motion velocity can be calculated from blurred images.On the other hand,the blurred image can also be restored using the obtained motion information.For solving a problem with small motion limitation on the general optical flow computation,a multiresolution optical flow algoritm based on MAP estimation is proposed. For restoring the blurred image ,an iteration algorithm and the obtained motion velocity are used.The experiment shows that the proposed approach for both motion velocity computation and blurred image restoration works well.     

Three-dimensional scene flow

Just as optical flow is the two-dimensional motion of points in an image, scene flow is the three-dimensional motion of points in the world. The fundamental difficulty with optical flow is that only the normal flow can be computed directly from the image measurements, without some form of smoothing or regularization. In this paper, we begin by showing that the same fundamental limitation applies to scene flow; however, many cameras are used to image the scene. There are then two choices when computing scene flow: 1) perform the regularization in the images or 2) perform the regularization on the surface of the object in the scene. In this paper, we choose to compute scene flow using regularization in the images. We describe three algorithms, the first two for computing scene flow from optical flows and the third for constraining scene structure from the inconsistencies in multiple optical flows.     

Generalized parallel-perspective stereo mosaics from airborne video

In this paper, we present a new method for automatically and efficiently generating stereoscopic mosaics by seamless registration of images collected by a video camera mounted on an airborne platform. Using a parallel-perspective representation, a pair of geometrically registered stereo mosaics can be precisely constructed under quite general motion. A novel parallel ray interpolation for stereo mosaicing (PRISM) approach is proposed to make stereo mosaics seamless in the presence of obvious motion parallax and for rather arbitrary scenes. Parallel-perspective stereo mosaics generated with the PRISM method have better depth resolution than perspective stereo due to the adaptive baseline geometry. Moreover, unlike previous results showing that parallel-perspective stereo has a constant depth error, we conclude that the depth estimation error of stereo mosaics is in fact a linear function of the absolute depths of a scene. Experimental results on long video sequences are given.     

Variational Recursive Joint Estimation of Dense Scene Structure and Camera Motion from Monocular High Speed Traffic Sequences

We present an approach to jointly estimating camera motion and dense structure of a static scene in terms of depth maps from monocular image sequences in driver-assistance scenarios. At each instant of time, only two consecutive frames are processed as input data of a joint estimator that fully exploits second-order information of the corresponding optimization problem and effectively copes with the non-convexity due to both the imaging geometry and the manifold of motion parameters. Additionally, carefully designed Gaussian approximations enable probabilistic inference based on locally varying confidence and globally varying sensitivity due to the epipolar geometry, with respect to the high-dimensional depth map estimation. Embedding the resulting joint estimator in an online recursive framework achieves a pronounced spatio-temporal filtering effect and robustness. We evaluate hundreds of images taken from a car moving at speed up to 100 km/h and being part of a publicly available benchmark data set. The results compare favorably with two alternative settings: stereo based scene reconstruction and camera motion estimation in batch mode using multiple frames. They, however, require a calibrated camera pair or storage for more than two frames, which is less attractive from a technical viewpoint than the proposed monocular and recursive approach. In addition to real data, a synthetic sequence is considered which provides reliable ground truth.     

基于场景对象注意与深度图融合的深度估计

现有单目深度估计算法主要从单幅图像中获取立体信息,存在相邻深度边缘细节模糊、明显的对象缺失问题。提出一种基于场景对象注意机制与加权深度图融合的单目深度估计算法。通过特征矩阵相乘的方式计算特征图任意两个位置之间的相似特征向量,以快速捕获长距离依赖关系,增强用于估计相似深度区域的上下文信息,从而解决自然场景中对象深度信息不完整的问题。基于多尺度特征图融合的优点,设计加权深度图融合模块,为具有不同深度信息的多视觉粒度的深度图赋予不同的权值并进行融合,融合后的深度图包含深度信息和丰富的场景对象信息,有效地解决细节模糊问题。在KITTI数据集上的实验结果表明,该算法对目标图像预估时σ<1.25的准确率为0.879,绝对相对误差、平方相对误差和对数均方根误差分别为0.110、0.765和0.185,预测得到的深度图具有更加完整的场景对象轮廓和精确的深度信息。     

任意视角的多视图立体匹配系统

为了得到高精度深度图,通过特征点提取与匹配、计算基础矩阵F、引导互匹配、捆集调整等一系列技术,求出每幅视图在同一空间坐标系下的高精度摄像机矩阵P.为任意角度的多个视图建立可扩展的主框架,不需要图像矫正.基于图像坐标和自动计算出的视差范围建立三维网络,为每条边分配适当的权值,把多视图立体匹配问题转化为网络最大流问题.算法保证了高准确率和平滑性,实验结果表明此方法适用于三维模型重建.     

外极面图像的运动遮挡模型和运动纹理方向检测算法

运动遮挡边界处的运动估计是一种困难的问题,外极面图像方法将运动估计转化为转迹线的检测,人造物体的轨迹线容易通过边缘跟踪的方法获得,但对于纹理复杂的自然景物,轨迹跟踪较为困难。     

基于深度学习的单目深度估计技术综述

场景的深度估计问题是计算机视觉领域中的经典问题之一,也是3维重建和图像合成等应用中的一个重要环节。基于深度学习的单目深度估计技术高速发展,各种网络结构相继提出。本文对基于深度学习的单目深度估计技术最新进展进行了综述,回顾了基于监督学习和基于无监督学习方法的发展历程。重点关注单目深度估计的优化思路及其在深度学习网络结构中的表现,将监督学习方法分为多尺度特征融合的方法、结合条件随机场(conditional random field,CRF)的方法、基于序数关系的方法、结合多元图像信息的方法和其他方法等5类;将无监督学习方法分为基于立体视觉的方法、基于运动恢复结构(structure from motion,SfM)的方法、结合对抗性网络的方法、基于序数关系的方法和结合不确定性的方法等5类。此外,还介绍了单目深度估计任务中常用的数据集和评价指标,并对目前基于深度学习的单目深度估计技术在精确度、泛化性、应用场景和无监督网络中不确定性研究等方面的现状和面临的挑战进行了讨论,为相关领域的研究人员提供一个比较全面的参考。