scispace - formally typeset
Search or ask a question

Showing papers presented at "German Conference on Pattern Recognition in 2016"


Book ChapterDOI
12 Sep 2016
TL;DR: This work presents an approach that transfers the style from one image (for example, a painting) to a whole video sequence, and makes use of recent advances in style transfer in still images and proposes new initializations and loss functions applicable to videos.
Abstract: In the past, manually re-drawing an image in a certain artistic style required a professional artist and a long time. Doing this for a video sequence single-handed was beyond imagination. Nowadays computers provide new possibilities. We present an approach that transfers the style from one image (for example, a painting) to a whole video sequence. We make use of recent advances in style transfer in still images and propose new initializations and loss functions applicable to videos. This allows us to generate consistent and stable stylized video sequences, even in cases with large motion and strong occlusion. We show that the proposed method clearly outperforms simpler baselines both qualitatively and quantitatively.

229 citations


Book ChapterDOI
12 Sep 2016
TL;DR: In this paper, a fully convolutional network (FCN) is used to predict semantic labels, depth and an instance-based encoding using each pixel's direction towards its corresponding instance center.
Abstract: Recent approaches for instance-aware semantic labeling have augmented convolutional neural networks (CNNs) with complex multi-task architectures or computationally expensive graphical models. We present a method that leverages a fully convolutional network (FCN) to predict semantic labels, depth and an instance-based encoding using each pixel’s direction towards its corresponding instance center. Subsequently, we apply low-level computer vision techniques to generate state-of-the-art instance segmentation on the street scene datasets KITTI and Cityscapes. Our approach outperforms existing works by a large margin and can additionally predict absolute distances of individual instances from a monocular image as well as a pixel-level semantic labeling.

135 citations


Book ChapterDOI
12 Sep 2016
TL;DR: This work presents a novel method for accurate and efficient upsampling of sparse depth data, guided by high-resolution imagery that determines globally consistent solutions and preserves fine details and sharp depth boundaries.
Abstract: We present a novel method for accurate and efficient upsampling of sparse depth data, guided by high-resolution imagery. Our approach goes beyond the use of intensity cues only and additionally exploits object boundary cues through structured edge detection and semantic scene labeling for guidance. Both cues are combined within a geodesic distance measure that allows for boundary-preserving depth interpolation while utilizing local context. We model the observed scene structure by locally planar elements and formulate the upsampling task as a global energy minimization problem. Our method determines globally consistent solutions and preserves fine details and sharp depth boundaries. In our experiments on several public datasets at different levels of application, we demonstrate superior performance of our approach over the state-of-the-art, even for very sparse measurements.

91 citations


Book ChapterDOI
12 Sep 2016
TL;DR: This paper builds on convolutional neural networks, which lead to significantly superior results compared with previous state-of-the-art on hand-crafted recognition pipelines, and shows how to further increase discrimination abilities of CNN activations by the Log-Euclidean framework on top of bilinear pooling.
Abstract: In this paper, we investigate how to predict attributes of chimpanzees such as identity, age, age group, and gender. We build on convolutional neural networks, which lead to significantly superior results compared with previous state-of-the-art on hand-crafted recognition pipelines. In addition, we show how to further increase discrimination abilities of CNN activations by the Log-Euclidean framework on top of bilinear pooling. We finally introduce two curated datasets consisting of chimpanzee faces with detailed meta-information to stimulate further research. Our results can serve as the foundation for automated large-scale animal monitoring and analysis.

78 citations


Book ChapterDOI
12 Sep 2016
TL;DR: This work proposes a novel approach for using compact shape manifolds of the shape within an object class for object segmentation, pose and shape estimation and demonstrates that the shape manifold alignment method yields improved results over the initial stereo reconstruction and object detection method in depth and pose accuracy.
Abstract: Estimating the pose and 3D shape of a large variety of instances within an object class from stereo images is a challenging problem, especially in realistic conditions such as urban street scenes. We propose a novel approach for using compact shape manifolds of the shape within an object class for object segmentation, pose and shape estimation. Our method first detects objects and estimates their pose coarsely in the stereo images using a state-of-the-art 3D object detection method. An energy minimization method then aligns shape and pose concurrently with the stereo reconstruction of the object. In experiments, we evaluate our approach for detection, pose and shape estimation of cars in real stereo images of urban street scenes. We demonstrate that our shape manifold alignment method yields improved results over the initial stereo reconstruction and object detection method in depth and pose accuracy.

78 citations


Book ChapterDOI
12 Sep 2016
TL;DR: A novel scale selection layer which extracts convolutional features at the scale which matches the corresponding reconstructed depth and frees the pixel-level classifier from the need to learn the laws of the perspective results in improved segmentation results.
Abstract: We propose an effective technique to address large scale variation in images taken from a moving car by cross-breeding deep learning with stereo reconstruction. Our main contribution is a novel scale selection layer which extracts convolutional features at the scale which matches the corresponding reconstructed depth. The recovered scale-invariant representation disentangles appearance from scale and frees the pixel-level classifier from the need to learn the laws of the perspective. This results in improved segmentation results due to more efficient exploitation of representation capacity and training data. We perform experiments on two challenging stereoscopic datasets (KITTI and Cityscapes) and report competitive class-level IoU performance.

65 citations


Book ChapterDOI
12 Sep 2016
TL;DR: This work proposes a convnet designed to perform non-maximum suppression of a given set of detections, and overcomes the intrinsic limitations of greedy NMS, obtaining better recall and precision.
Abstract: Non-maximum suppression (NMS) is used in virtually all state-of-the-art object detection pipelines. While essential object detection ingredients such as features, classifiers, and proposal methods have been extensively researched surprisingly little work has aimed to systematically address NMS. The de-facto standard for NMS is based on greedy clustering with a fixed distance threshold, which forces to trade-off recall versus precision. We propose a convnet designed to perform NMS of a given set of detections. We report experiments on a synthetic setup, crowded pedestrian scenes, and for general person detection. Our approach overcomes the intrinsic limitations of greedy NMS, obtaining better recall and precision.

54 citations


Book ChapterDOI
12 Sep 2016
TL;DR: An active learning and discovery approach which can deal with huge collections of unlabeled real-world data based on the expected model output change principle and overcomes previous scalability issues is presented.
Abstract: Incremental learning of visual concepts is one step towards reaching human capabilities beyond closed-world assumptions. Besides recent progress, it remains one of the fundamental challenges in computer vision and machine learning. Along that path, techniques are needed which allow for actively selecting informative examples from a huge pool of unlabeled images to be annotated by application experts. Whereas a manifold of active learning techniques exists, they commonly suffer from one of two drawbacks: (i) either they do not work reliably on challenging real-world data or (ii) they are kernel-based and not scalable with the magnitudes of data current vision applications need to deal with. Therefore, we present an active learning and discovery approach which can deal with huge collections of unlabeled real-world data. Our approach is based on the expected model output change principle and overcomes previous scalability issues. We present experiments on the large-scale MS-COCO dataset and on a dataset provided by biodiversity researchers. Obtained results reveal that our technique clearly improves accuracy after just a few annotations. At the same time, it outperforms previous active learning approaches in academic and real-world scenarios.

20 citations


Book ChapterDOI
12 Sep 2016
TL;DR: In this paper, the authors focus on the problem of explaining predictions of psychological attributes such as attractiveness, happiness, confidence and intelligence from face photographs using deep neural networks and apply transfer learning with two base models to avoid overfitting.
Abstract: This paper focuses on the problem of explaining predictions of psychological attributes such as attractiveness, happiness, confidence and intelligence from face photographs using deep neural networks. Since psychological attribute datasets typically suffer from small sample sizes, we apply transfer learning with two base models to avoid overfitting. These models were trained on an age and gender prediction task, respectively. Using a novel explanation method we extract heatmaps that highlight the parts of the image most responsible for the prediction. We further observe that the explanation method provides important insights into the nature of features of the base model, which allow one to assess the aptitude of the base model for a given transfer learning task. Finally, we observe that the multiclass model is more feature rich than its binary counterpart. The experimental evaluation is performed on the 2222 images from the 10k US faces dataset containing psychological attribute labels as well as on a subset of KDEF images.

20 citations


Book ChapterDOI
12 Sep 2016
TL;DR: This work proposes an original two-step method for fast and performant optical flow estimation from stereo vision, which achieves framerate processing on images of realistic size, and provides results comparable or better than methods having computation times one or two orders of magnitude higher.
Abstract: Estimating the optical flow robustly in real-time is still a challenging issue as revealed by current KITTI benchmarks. We propose an original two-step method for fast and performant optical flow estimation from stereo vision. The first step is the prediction of the flow due to the ego-motion, efficiently conducted by stereo-matching and visual odometry. The correction step estimates the motion of mobile objects. Algorithmic choices are justified by empirical studies on real datasets. Our method achieves framerate processing on images of realistic size, and provides results comparable or better than methods having computation times one or two orders of magnitude higher.

16 citations


Book ChapterDOI
12 Sep 2016
TL;DR: A simple and highly efficient acceleration strategy is introduced, leading to so-called Fast Semi-Iterative (FSI) schemes that extrapolate the basic solver iteration with the previous iterate to derive suitable extrapolation parameters.
Abstract: Many tasks in image processing and computer vision are modelled by diffusion processes, variational formulations, or constrained optimisation problems. Basic iterative solvers such as explicit schemes, Richardson iterations, or projected gradient descent methods are simple to implement and well-suited for parallel computing. However, their efficiency suffers from severe step size restrictions. As a remedy we introduce a simple and highly efficient acceleration strategy, leading to so-called Fast Semi-Iterative (FSI) schemes that extrapolate the basic solver iteration with the previous iterate. To derive suitable extrapolation parameters, we establish a recursion relation that connects box filtering with an explicit scheme for 1D homogeneous diffusion. FSI schemes avoid the main drawbacks of recent Fast Explicit Diffusion (FED) and Fast Jacobi techniques, and they have an interesting connection to the heavy ball method in optimisation. Our experiments show their benefits for anisotropic diffusion inpainting, nonsmooth regularisation, and Nesterov’s worst case problems for convex and strongly convex optimisation.

Book ChapterDOI
12 Sep 2016
TL;DR: A remarkable feature of the proposed method is the ability to correct some erroneous labels, when computer generated initial labels are considered, which allows it to improve state-of-the-art methods for motion segmentation in videos by 5–10 % with respect to the F-measure (Dice score).
Abstract: Recent advances on convex relaxation methods allow for a flexible formulation of many interactive multi-label segmentation methods. The building blocks are a likelihood specified for each pixel and each label, and a penalty for the boundary length of each segment. While many sophisticated likelihood estimations based on various statistical measures have been investigated, the boundary length is usually measured in a metric induced by simple image gradients. We show that complementing these methods with recent advances of edge detectors yields an immense quality improvement. A remarkable feature of the proposed method is the ability to correct some erroneous labels, when computer generated initial labels are considered. This allows us to improve state-of-the-art methods for motion segmentation in videos by 5–10 % with respect to the F-measure (Dice score).

Book ChapterDOI
12 Sep 2016
TL;DR: A novel method for the detection of vibrations caused by trains in an optical fiber buried nearby the railway track using optical time-domain reflectometry vibrations in the ground caused by different sources can be detected with high accuracy in time and space.
Abstract: We propose a novel method for the detection of vibrations caused by trains in an optical fiber buried nearby the railway track. Using optical time-domain reflectometry vibrations in the ground caused by different sources can be detected with high accuracy in time and space. While several algorithms have been proposed in the literature for train tracking using OTDR signals they have not been tested on longer recordings. The presented method learns the characteristic pattern in the Fourier domain using a support vector machine (SVM) and it becomes more robust to any kind of noise and artifacts in the signal. The point-based causal train tracking has two stages to minimize the influence of false classifications of the vibration detection. Our technical contribution is the evaluation of the presented algorithm based on two hour long recording and demonstration of open problems for commercial usage.

Book ChapterDOI
12 Sep 2016
TL;DR: A regularization procedure based on the nonlinear Landweber method is introduced for the stable determination of the source location of density functions evolved within a nonlinear reaction-diffusion model for brain tumors.
Abstract: We propose a mathematically well-founded approach for locating the source (initial state) of density functions evolved within a nonlinear reaction-diffusion model. The reconstruction of the initial source is an ill-posed inverse problem since the solution is highly unstable with respect to measurement noise. To address this instability problem, we introduce a regularization procedure based on the nonlinear Landweber method for the stable determination of the source location. This amounts to solving a sequence of well-posed forward reaction-diffusion problems. The developed framework is general, and as a special instance we consider the problem of source localization of brain tumors. We show numerically that the source of the initial densities of tumor cells are reconstructed well on both imaging data consisting of simple and complex geometric structures.

Book ChapterDOI
12 Sep 2016
TL;DR: The focus here is on estimating the ego-motion rather than a detailed reconstruction of the scene, and a slim method is proposed that is within the top ranks of the KITTI benchmark without using any filtering method like bundle adjustment or Kalman filtering.
Abstract: Visual Odometry is one of the key technology for navigating and perceiving the environment of an autonomous vehicle. Within the last ten years, a common sense has been established on how to implement high precision and robust systems. This paper goes one step back by avoiding temporal filtering and relying exclusively on pure measurements that have been carefully selected. The focus here is on estimating the ego-motion rather than a detailed reconstruction of the scene. Different approaches for selecting proper 3D-flows (scene flows) are compared and discussed. The ego-motion is computed by a standard P6P-approach encapsulated in a RANSAC environment. Finally, a slim method is proposed that is within the top ranks of the KITTI benchmark without using any filtering method like bundle adjustment or Kalman filtering.

Book ChapterDOI
12 Sep 2016
TL;DR: This work reduces the number of necessary dictionary atoms, improving descriptive quality of each and reducing time complexity by an order of magnitude, and introduces a way to explicitly handle occlusions, which is the main drawback in the previous work.
Abstract: Disparity estimation for multi-layered light fields can robustly be performed with a statistical analysis of sparse light field coding coefficients [7]. The key idea is to explain each epipolar plane image patch with a dictionary composed of atoms with known disparity values. We significantly improve upon their approach in two ways. First, we reduce the number of necessary dictionary atoms, improving descriptive quality of each and reducing time complexity by an order of magnitude. Second, we introduce a way to explicitly handle occlusions, which is the main drawback in the previous work. Experiments demonstrate that we thus achieve substantially better results on both Lambertian as well as multi-layered scenes.

Book ChapterDOI
12 Sep 2016
TL;DR: This work addresses the problem of semantic segmentation of objects in weakly supervised setting, when only image-wide labels are available, by describing an image with a set of pre-trained convolutional features and embeding this set into a Fisher vector.
Abstract: We address the problem of semantic segmentation of objects in weakly supervised setting, when only image-wide labels are available. We describe an image with a set of pre-trained convolutional features and embed this set into a Fisher vector. We apply the learned image classifier on the set of all image regions and propagate the region scores back to the pixels. Compared to the alternatives the proposed method is simple, fast in inference, and especially in training. The method displays very good performance of on two standard semantic segmentation benchmarks.

Book ChapterDOI
12 Sep 2016
TL;DR: A generative model of a light field that is fully parametrized by its corresponding depth map is proposed, which allows for the integration of powerful regularization techniques such as a non-local means prior, facilitating accurate depth map estimation.
Abstract: Light field photography captures rich structural information that may facilitate a number of traditional image processing and computer vision tasks. A crucial ingredient in such endeavors is accurate depth recovery. We present a novel framework that allows the recovery of a high quality continuous depth map from light field data. To this end we propose a generative model of a light field that is fully parametrized by its corresponding depth map. The model allows for the integration of powerful regularization techniques such as a non-local means prior, facilitating accurate depth map estimation. Comparisons with previous methods show that we are able to recover faithful depth maps with much finer details. In a number of challenging real-world examples we demonstrate both the effectiveness and robustness of our approach.


Book ChapterDOI
12 Sep 2016
TL;DR: A non-convex variational approach to non-binary discrete tomography which combines non-local projection constraints with a continuous convex relaxation of the multilabeling problem is presented.
Abstract: We present a non-convex variational approach to non-binary discrete tomography which combines non-local projection constraints with a continuous convex relaxation of the multilabeling problem. Minimizing this non-convex energy is achieved by a fixed point iteration which amounts to solving a sequence of convex problems, with guaranteed convergence to a critical point. A competitive numerical evaluation using standard test-datasets demonstrates a significantly improved reconstruction quality for noisy measurements from a small number of projections.

Book ChapterDOI
12 Sep 2016
TL;DR: This paper introduces a fully automatic fast method for depth estimation from a single plenoptic image running a RANSAC-like algorithm for feature matching and achieves an accuracy similar to the state-of-the-art in considerable less time.
Abstract: Light field cameras capture a scene’s multi-directional light field with one image, allowing the estimation of depth. In this paper, we introduce a fully automatic fast method for depth estimation from a single plenoptic image running a RANSAC-like algorithm for feature matching. The novelty about our approach is the global method to back project correspondences found using photometric similarity to obtain a 3D virtual point cloud. We then use lenses with different focal-lengths in a multiple depth map refining phase and their reprojection to the image plane, generating an accurate depth map per micro lens. Tests with simulations and real images are presented and show a good trade-off between computation time and accuracy of the method presented. Our method achieves an accuracy similar to the state-of-the-art in considerable less time (speedups of around 3 times).

Book ChapterDOI
12 Sep 2016
TL;DR: In this article, the authors characterize the tightest convex extension of the objective function, given by the Legendre-Fenchel biconjugate, and derive a closed form.
Abstract: Regularized empirical risk minimization with constrained labels (in contrast to fixed labels) is a remarkably general abstraction of learning. For common loss and regularization functions, this optimization problem assumes the form of a mixed integer program (MIP) whose objective function is non-convex. In this form, the problem is resistant to standard optimization techniques. We construct MIPs with the same solutions whose objective functions are convex. Specifically, we characterize the tightest convex extension of the objective function, given by the Legendre-Fenchel biconjugate. Computing values of this tightest convex extension is NP-hard. However, by applying our characterization to every function in an additive decomposition of the objective function, we obtain a class of looser convex extensions that can be computed efficiently. For some decompositions, common loss and regularization functions, we derive a closed form.

Book ChapterDOI
12 Sep 2016
TL;DR: This paper proposes a method for automated detection and segmentation of immunostained cell nuclei in ultramicroscopy images using interactive learning and voxel classification and performs the splitting process for each cluster using a multi-step watershed approach.
Abstract: Detection, segmentation, and quantification of individual cell nuclei is a standard task in biomedical applications. Due to the increasing volume of acquired image data, it is not possible to rely on manual labeling and object counting. Instead, automated image processing methods have to be applied. Especially in three-dimensional data, one of the major challenges is the separation of touching cell nuclei in densely packed clusters. In this paper, we propose a method for automated detection and segmentation of immunostained cell nuclei in ultramicroscopy images. Our algorithm utilizes interactive learning and voxel classification to obtain a foreground segmentation and subsequently performs the splitting process for each cluster using a multi-step watershed approach. We have evaluated our results using reference images manually labeled by domain experts and compare our approach to state-of-the art methods.

Book ChapterDOI
12 Sep 2016
TL;DR: An algorithm for Contiguous PAtch Segmentation (CPAS) in 3D pointclouds is proposed that is robust, scalable and provides a more complete description by simultaneously detecting contiguous patches as well as delineating object boundaries.
Abstract: An algorithm for Contiguous PAtch Segmentation (CPAS) in 3D pointclouds is proposed. In contrast to current state-of-the-art algorithms, CPAS is robust, scalable and provides a more complete description by simultaneously detecting contiguous patches as well as delineating object boundaries. Our algorithm uses a voxel grid to divide the scene into non-overlapping voxels within which clipped planes are fitted with RANSAC. Using a Dirichlet process mixture (DPM) model of Gaussians and connected component analysis, voxels are clustered into contiguous regions. Finally, we use importance sampling on the convex-hull of each region to obtain the underlying patch and object boundary estimates. For urban scenes, the segmentation represents building walls, ground and roof elements (Fig. 1). We demonstrate the robustness of CPAS using data sets from both image matching and raw LiDAR scans.

Book ChapterDOI
12 Sep 2016
TL;DR: This paper analyses the applicability and performance of Convolutional Neural Networks to localise and segment anatomical structures in medical volumes under clinically realistic constraints: small amount of available training data, need of a short processing time and limited computational resources.
Abstract: This paper analyses the applicability and performance of Convolutional Neural Networks (CNN) to localise and segment anatomical structures in medical volumes under clinically realistic constraints: small amount of available training data, the need of a short processing time and limited computational resources. Our segmentation approach employs CNNs for simultaneous classification and feature extraction. A Hough voting strategy has been developed in order to automatically localise and segment the anatomy of interest. Our results show (i) improved robustness, due to the inclusion of prior shape knowledge, (ii) highly accurate segmentation even when only small datasets are available during training, (iii) speed and computational requirements that match those that are usually present in clinical settings.

Book ChapterDOI
12 Sep 2016
TL;DR: The combination of the edge detection and the line detection to a single step provides a more robust and more reliable detection of larger line segments.
Abstract: This article presents a method to detect lines in fisheye and distorted perspective images. The detection is performed with subpixel accuracy. By detecting lines in the original images without warping the image with a reverse distortion, the detection accuracy can be noticeably improved. The combination of the edge detection and the line detection to a single step provides a more robust and more reliable detection of larger line segments.

Book ChapterDOI
12 Sep 2016
TL;DR: A novel co-segmentation approach that constructs a part-based object representation comprised of shape appearance models of individual parts and isometric spatial relations between the parts and a compact Conditional Random Field is proposed.
Abstract: The practical use of the latest methods for supervised 3D shape co-segmentation is limited by the requirement of diverse training data and a watertight mesh representation. Driven by practical considerations, we assume only one reference shape to be available and the query shape to be provided as a partially visible point cloud. We propose a novel co-segmentation approach that constructs a part-based object representation comprised of shape appearance models of individual parts and isometric spatial relations between the parts. The partial query shape is pre-segmented using planar cuts, and the segments accompanied by the learned representation induce a compact Conditional Random Field (CRF). CRF inference is performed efficiently by \(A^*\)-search with global optimality guarantees. A comparative evaluation with two baselines on partial views generated from the Labelled Princeton Segmentation Benchmark and point clouds recorded with an RGB-D sensor demonstrate superiority of the proposed approach both in accuracy and efficiency.

Book ChapterDOI
12 Sep 2016
TL;DR: This work leverages prior work on casting the computation of a 2.5D depth map as a labeling problem and shows that this formulation has great potential as an intermediate representation in the context of building facade reconstruction.
Abstract: Multi-View Stereo offers an affordable and flexible method for the acquisition of 3D point clouds. However, these point clouds are prone to errors and missing regions. In addition, an abstraction in the form of a simple mesh capturing the essence of the surface is usually preferred over the raw point cloud measurement. We present a fully automatic pipeline that computes such a mesh from the noisy point cloud of a building facade. We leverage prior work on casting the computation of a 2.5D depth map as a labeling problem and show that this formulation has great potential as an intermediate representation in the context of building facade reconstruction.

Book ChapterDOI
12 Sep 2016
TL;DR: This paper presents a method to generate a meaningful confidence measurement during online real-time egomotion estimation of a vehicle using a monocular camera and shows that this confidence measurement gives reliable results and can be used to filter the egomotions estimation using a Kalman filter.
Abstract: This paper presents a method to generate a meaningful confidence measurement during online real-time egomotion estimation of a vehicle using a monocular camera. This confidence measurement should give the information whether the signal fulfills a certain accuracy range in all parameters or not. For that reason features from an optical flow field incorporating the egomotion error are determined and a confidence measurement is learned using ground truth egomotion data that we obtain from an offline bundle adjustment before. We show that our confidence measurement gives reliable results and can further be used to filter the egomotion estimation using a Kalman filter. Incorporating the knowledge of the egomotion accuracy determined by the confidence we are able to update the confidence measure for the filtered results. This leads to an improved system availability.

Book ChapterDOI
12 Sep 2016
TL;DR: This study presents several experiments, which show that motion fields are more accurate when computed using image derivatives evaluated by regularized variational differentiation than with conventional averaging of finite differences.
Abstract: The purpose of this study is to investigate image differentiation by a boundary preserving variational method. The method minimizes a functional composed of an anti-differentiation data discrepancy term and an \(L^1\) regularization term. For each partial derivative of the image, the anti-differentiation term biases the minimizer toward a function which integrates to the image up to an additive constant, while the regularization term biases it toward a function smooth everywhere except across image edges. A discretization of the functional Euler-Lagrange equations gives a large scale system of nonlinear equations that, however, is sparse, and “almost” linear, which directs to a resolution by successive linear approximations. The method is investigated in two important computer vision problems, namely optical flow and scene flow estimation, where image differentiation is used and ordinarily done by local averaging of finite image differences. We present several experiments, which show that motion fields are more accurate when computed using image derivatives evaluated by regularized variational differentiation than with conventional averaging of finite differences.