Highly accurate optical flow estimation on superpixel tree

Formulated as a pixel-labeling problem, optical flow estimation using efficient edge-aware filtering has shown great success recently. However, the typical challenge that restricts the range of applicability of this method is the computational complexity mainly caused by the testing of every hypothetical label in the whole label space, which is usually large in an optical flow estimation. In this paper, we present an effective and efficient two-level filter-based optical flow algorithm connected by an accurate non-local matching. With the key observation that the optical flow of the pixels from the same compact superpixels is highly coherent, we propose a novel superpixel tree representation of an image to obtain an accurate superpixel flow. We find that if filtered separately, the candidate label space of the pixels from each superpixel is drastically reduced with the known superpixel flow. We also suggest a refined label selection strategy that is more accurate than the usual winner-takes-all manner. The proposed method, called Highly Accurate flow on Superpixel Tree (HastFlow) is validated on Middlebury and MPI-Sintel, and outperforms all filter-based methods both in accuracy and efficiency.     

Motion detail preserving optical flow estimation

A common problem of optical flow estimation in the multiscale variational framework is that fine motion structures cannot always be correctly estimated, especially for regions with significant and abrupt displacement variation. A novel extended coarse-to-fine (EC2F) refinement framework is introduced in this paper to address this issue, which reduces the reliance of flow estimates on their initial values propagated from the coarse level and enables recovering many motion details in each scale. The contribution of this paper also includes adaptation of the objective function to handle outliers and development of a new optimization procedure. The effectiveness of our algorithm is demonstrated by Middlebury optical flow benchmarkmarking and by experiments on challenging examples that involve large-displacement motion.     

光照变化条件下的光流估计

目的 为了提高光流法在处理光照变化和大位移方面的稳健性。提出一种结合结构纹理分解预处理和加权中值滤波的光流场模型。方法 该方法数据项采用灰度守恒假设和梯度守恒假设相结合、局部约束与全局约束相结合的思想。同时采用结构纹理分解、加权中值滤波、金字塔结构等高效的光流估计技术,进一步增强了光流算法的精确性与实用性。结果 分别通过Middlebury光流数据库图像和真实场景图像对提出的光流估计算法进行了大量实验验证。实验结果表明,改进的光流估计法在处理光照变化方面表现不错,不仅获得稠密的光流场,而且提高了光流场准确提取目标边缘的能力。结论 和传统光流方法相比,所提方法在光照变化情况下能获得更加理想的结果,降低了实际场景中光线变化的干扰,能更好地适用于实际场景中。     

Median radial basis function neural network

Radial basis functions (RBFs) consist of a two-layer neural network, where each hidden unit implements a kernel function. Each kernel is associated with an activation region from the input space and its output is fed to an output unit. In order to find the parameters of a neural network which embeds this structure we take into consideration two different statistical approaches. The first approach uses classical estimation in the learning stage and it is based on the learning vector quantization algorithm and its second-order statistics extension. After the presentation of this approach, we introduce the median radial basis function (MRBF) algorithm based on robust estimation of the hidden unit parameters. The proposed algorithm employs the marginal median for kernel location estimation and the median of the absolute deviations for the scale parameter estimation. A histogram-based fast implementation is provided for the MRBF algorithm. The theoretical performance of the two training algorithms is comparatively evaluated when estimating the network weights. The network is applied in pattern classification problems and in optical flow segmentation.     

基于运动优化语义分割的变分光流计算方法

针对光照变化和大位移运动等复杂场景下图像序列变分光流计算的边缘模糊与过度分割问题,文中提出基于运动优化语义分割的变分光流计算方法.首先,根据图像局部区域的去均值归一化匹配模型,构建变分光流计算能量泛函.然后,利用去均值归一化互相关光流估计结果,获取图像运动边界信息,优化语义分割,设计运动约束语义分割的变分光流计算模型....     

A Variational Aggregation Framework for Patch-Based Optical Flow Estimation

We propose a variational aggregation method for optical flow estimation. It consists of a two-step framework, first estimating a collection of parametric motion models to generate motion candidates, and then reconstructing a global dense motion field. The aggregation step is designed as a motion reconstruction problem from spatially varying sets of motion candidates given by parametric motion models. Our method is designed to capture large displacements in a variational framework without requiring any coarse-to-fine strategy. We handle occlusion with a motion inpainting approach in the candidates computation step. By performing parametric motion estimation, we combine the robustness to noise of local parametric methods with the accuracy yielded by global regularization. We demonstrate the performance of our aggregation approach by comparing it to standard variational methods and a discrete aggregation approach on the Middlebury and MPI Sintel datasets.     

Improving the robustness of variational optical flow through tensor voting

Differential optical flow methods allow the estimation of optical flow fields based on the first-order and even higher-order spatio-temporal derivatives (gradients) of sequences of input images. If the input images are noisy, for instance because of the limited quality of the capturing devices or due to poor illumination conditions, the use of partial derivatives will amplify that noise and thus end up affecting the accuracy of the computed flow fields. The typical approach in order to reduce that noise consists of smoothing the required gradient images with Gaussian filters, for instance by applying structure tensors. However, that filtering is isotropic and tends to blur the discontinuities that may be present in the original images, thus likely leading to an undesired loss of accuracy in the resulting flow fields. This paper proposes the use of tensor voting as an alternative to Gaussian filtering, and shows that the discontinuity preserving capabilities of the former yield more robust and accurate results. In particular, a state-of-the-art variational optical flow method has been adapted in order to utilize a tensor voting filtering approach. The proposed technique has been tested upon different datasets of both synthetic and real image sequences, and compared to both well known and state-of-the-art differential optical flow methods.     

Fast motion estimation for field sequential imaging: Survey and benchmark

Field sequential (FS) imaging comprises image acquisition systems that capture image channels in temporal sequence in order to provide the final image. A classical application is multispectral imaging. In case of dynamic scenes, the sequential nature of the acquisition imposes motion artifacts, i.e., spatially misaligned images channels. Compensating motion artifacts for this kind of imagery is non-trivial, as common methods for motion estimation rely on the intensity consistency constraint that is violated in FS imaging.This paper surveys approaches to motion compensation in the context of FS imaging. We focus on accuracy in handling intensity inconsistent data and, secondarily, speed, as FS imaging is commonly done in real-time. We introduce a conceptual classification for algorithmic approaches for motion estimation for FS imagery and discuss known and modified approaches to tackle the intensity inconsistencies between adjacent image channels using image transformation and intensity correction methods. As result, we get a set of 379 variants of motion estimation methods applicable to FS data streams. We evaluate these methods using our benchmark database, which comprises data sets from the Middlebury and the MPI Sintel databases, modified to emulate FS imagery, as well as additionally captured multispectral short wave infrared (SWIR) and sRGB image sequences, as well as simulated Time-of-Flight (ToF) image sequences that consist of four channels (called phase images). In order to quantify the motion estimation techniques, we use a ranking scheme similar to Middlebury and combine it with a run-time evaluation.     

基于多尺度变形卷积的特征金字塔光流计算方法

针对现有深度学习光流计算方法的运动边缘模糊问题,提出了一种基于多尺度变形卷积的特征金字塔光流计算方法.首先,构造基于多尺度变形卷积的特征提取模型,显著提高图像边缘区域特征提取的准确性;然后,将多尺度变形卷积特征提取模型与特征金字塔光流计算网络耦合,提出一种基于多尺度变形卷积的特征金字塔光流计算模型;最后,设计一种结合图像与运动边缘约束的混合损失函数,通过指导模型学习更加精准的边缘信息,克服了光流计算运动边缘模糊问题.分别采用MPI-Sintel和KITTI2015测试图像集对该方法与代表性的深度学习光流计算方法进行综合对比分析.实验结果表明,该方法具有更高的光流计算精度,有效解决了光流计算的边缘模糊问题.     

基于深度匹配的由稀疏到稠密大位移运动光流估计

针对非刚性大位移运动场景的光流计算准确性与鲁棒性问题,提出一种基于深度匹配的由稀疏到稠密大位移运动光流估计方法.首先利用深度匹配模型计算图像序列相邻帧的初始稀疏运动场;其次采用网格化邻域支持优化模型筛选具有较高置信度的图像网格和匹配像素点,获得鲁棒的稀疏运动场;然后对稀疏运动场进行边缘保护稠密插值,并设计全局能量泛函优化求解稠密光流场.最后分别利用MPI-Sintel和KITTI数据库提供的测试图像集对本文方法和Classic+NL,DeepFlow, EpicFlow以及FlowNetS等变分模型、匹配策略和深度学习光流计算方法进行综合对比与分析,实验结果表明本文方法相对于其他方法具有更高的光流计算精度,尤其在非刚性大位移和运动遮挡区域具有更好的鲁棒性与可靠性.     

结合改进密集模块深度估计网络和多视几何的视觉里程计

以多视图几何原理为基础,有效结合卷积神经网络进行图像深度估计和匹配筛选,构造无监督单目视觉里程计方法.针对主流深度估计网络易丢失图像浅层特征的问题,构造一种基于改进密集模块的深度估计网络,有效地聚合浅层特征,提升图像深度估计精度.里程计利用深度估计网络精确预测单目图像深度,利用光流网络获得双向光流,通过前后光流一致性原则筛选高质量匹配.利用多视图几何原理和优化方式求解获得初始位姿和计算深度,并通过特定的尺度对齐原则得到全局尺度一致的6自由度位姿.同时,为了提高网络对场景细节和弱纹理区域的学习能力,将基于特征图合成的特征度量损失结合到网络损失函数中.在KITTI Odometry数据集上进行实验验证,不同阈值下的深度估计取得了85.9%、95.8%、97.2%的准确率.在09和10序列上进行里程计评估,绝对轨迹误差在0.007 m.实验结果验证了所提出方法的有效性和准确性,表明其在深度估计和视觉里程计任务上的性能优于现有方法.     

Robust glottal source estimation based on joint source-filter model optimization

This paper describes a robust glottal source estimation method based on a joint source-filter separation technique. In this method, the Liljencrants-Fant (LF) model, which models the glottal flow derivative, is integrated into a time-varying ARX speech production model. These two models are estimated in a joint optimization procedure, in which a Kalman filtering process is embedded for adaptively identifying the vocal tract parameters. Since the formulated joint estimation problem is a multiparameter nonlinear optimization procedure, we separate the optimization procedure into two passes. The first pass initializes the glottal source and vocal tract models by solving a quasi-convex approximate optimization problem. Having robust initial values, the joint estimation procedure determines the accuracy of model estimation implemented with a trust-region descent optimization algorithm. Experiments with synthetic and real voice signals show that the proposed method is a robust glottal source parameter estimation method with a high degree of accuracy.     

深度光流估计方法研究进展

结合深度学习模型实现光流端到端的计算是当前计算机视觉领域的一个研究热点.文中对基于深度学习的光流估计方法进行总结和梳理.首先,介绍了光流的起源与定义;其次,总结了现有的数据集合和评价指标;最重要的是,着重从3个方面回顾了深度光流估计方法,包括有监督的深度光流估计方法、无监督的深度光流估计方法以及对现有光流估计方法的性能...     

Plug-and-Play Based Optimization Algorithm for New Crime Density Estimation

Different from a general density estimation, the crime density estimation usually has one important factor: the geographical constraint. In this paper, a new crime density estimation model is formulated, in which the regions where crime is impossible to happen, such as mountains and lakes, are excluded. To further optimize the estimation method, a learning-based algorithm, named Plug-and-Play, is implanted into the augmented Lagrangian scheme, which involves an off-the-shelf filtering operator. Different selections of the filtering operator make the algorithm correspond to several classical estimation models. Therefore, the proposed Plug-and-Play optimization based estimation algorithm can be regarded as the extended version and general form of several classical methods. In the experiment part, synthetic examples with different invalid regions and samples of various distributions are first tested. Then under complex geographic constraints, we apply the proposed method with a real crime dataset to recover the density estimation. The state-of-the-art results show the feasibility of the proposed model.

     

一种改进的全局光流估计方法

基于光流基本约束和平滑性约束条件的Horn-Schunck光流场佑计算法是图像运动估计的重要方法。但是,该方法存在在梯度值较小处运动参数估计不准确的问题;同时,现有的改进方法由于步及到可调参数的人工选取,并在阈值设置过高时容易在运动目标区域产生空洞,限制了光流法的应用。对光流基本约束项的权函数加以改进,给出了两种改进的光流估计算法。实验结果表明,改进算法能够在权函数阂值设置过高时降低对可靠光流的抑制,提高了算法的自适应性,为运动目标检测跟踪提供了有力条件。     

Test-Time Optimization for Video Depth Estimation Using Pseudo Reference Depth

In this paper, we propose a learning-based test-time optimization approach for reconstructing geometrically consistent depth maps from a monocular video. Specifically, we optimize an existing single image depth estimation network on the test example at hand. We do so by introducing pseudo reference depth maps which are computed based on the observation that the optical flow displacement for an image pair should be consistent with the displacement obtained by depth-reprojection. Additionally, we discard inaccurate pseudo reference depth maps using a simple median strategy and propose a way to compute a confidence map for the reference depth. We use our pseudo reference depth and the confidence map to formulate a loss function for performing the test-time optimization in an efficient and effective manner. We compare our approach against the state-of-the-art methods on various scenes both visually and numerically. Our approach is on average 2.5× faster than the state of the art and produces depth maps with higher quality.     

Mutual-Structure for Joint Filtering

Previous joint/guided filters directly transfer structural information from the reference to the target image. In this paper, we analyze the major drawback—that is, there may be completely different edges in the two images. Simply considering all patterns could introduce significant errors. To address this issue, we propose the concept of mutual-structure, which refers to the structural information that is contained in both images and thus can be safely enhanced by joint filtering. We also use an untraditional objective function that can be efficiently optimized to yield mutual structure. Our method results in important edge preserving property, which greatly benefits depth completion, optical flow estimation, image enhancement, stereo matching, to name a few.     

基于修正积分卡尔曼粒子滤波的自适应目标跟踪算法

针对当前粒子滤波权值退化问题以及精度与时耗的矛盾,提出了一种新的高精度自适应粒子滤波算法。该算法综合考虑优选建议分布函数和重采样两种并行改进滤波性能的方法:首先,在积分卡尔曼滤波(QKF)的基础上引入修正因子,通过修正的积分卡尔曼滤波(PQKF)产生优选的建议分布函数,较好地克服了粒子退化现象,在提高滤波精度的同时降低了运算量;在重采样阶段,通过引入系统估计和预测提供的新息差值在线自适应调整采样粒子数,较好地保证了粒子采样的高效性和算法的实时性。实验表明,新算法具有高精度、低时耗的优点,是一种高精度自适应粒子滤波算法。     

Over-Parameterized Variational Optical Flow

A novel optical flow estimation process based on a spatio-temporal model with varying coefficients multiplying a set of basis functions at each pixel is introduced. Previous optical flow estimation methodologies did not use such an over parameterized representation of the flow field as the problem is ill-posed even without introducing any additional parameters: Neighborhood based methods of the Lucas–Kanade type determine the flow at each pixel by constraining the flow to be described by a few parameters in small neighborhoods. Modern variational methods represent the optic flow directly via the flow field components at each pixel. The benefit of over-parametrization becomes evident in the smoothness term, which instead of directly penalizing for changes in the optic flow, accumulates a cost of deviating from the assumed optic flow model. Our proposed method is very general and the classical variational optical flow techniques are special cases of it, when used in conjunction with constant basis functions. Experimental results with the novel flow estimation process yield significant improvements with respect to the best results published so far.