Search

In the context of urban 3D mapping, the 3D modelling step is a crucial operation, which can be very error prone, particularly when high fidelity and accuracy are required. While automatic reconstruction tools are way faster and less expensive than manual ones, these last are less sensible to the data defects.To combine all this complementary advantages, we propose a semi-automatic approach, where human operators will manually correct the result of an automatic reconstruction tool.While the automated reconstruction tool will be taken on the shelf, we have developed our own polyhedral modeler tool to make the correction step. In this article, we present this polyhedral modeler, which is based on face shifting, edge flipping and automatic topological events resolution.

Recent advancements in 3D data acquisition and processing have enabled high-fidelity urban modeling. Yet, production of structured 3D models in standards like CityGML remain complex, resource-intensive, and difficult to automate. This paper introduces a low-cost alternative that we call “structured mesh model” designed to cover many applications of structured 3D models at a lower cost. It relies on integrating geometric simplification with segmentation alignment to produce a lightweight, unified mesh representation. Using an edge-collapse algorithm, our method combines geometry from an existing mesh with labeled point cloud data to create a continuous mesh with edges aligned to segmentation boundaries, preserving both geometric fidelity and semantic clarity. The resulting structured mesh efficiently reduces memory requirements while maintaining accuracy, offering a practical solution for simulations and urban analyses that require structured 3D data.

Textured meshes are becoming an increasingly popular representation combining the 3D geometry and radiometry of real scenes. However, semantic segmentation algorithms for urban mesh have been little investigated and do not exploit all radiometric information. To address this problem, we adopt an approach consisting in sampling a point cloud from the textured mesh, then using a point cloud semantic segmentation algorithm on this cloud, and finally using the obtained semantic to segment the initial mesh. In this paper, we study the influence of different parameters such as the sampling method, the density of the extracted cloud, the features selected (color, normal, elevation) as well as the number of points used at each training period. Our result outperforms the state-of-the-art on the SUM dataset, earning about 4 points in OA and 18 points in mIoU.

Combining image and LiDAR draws increasing interest in surface reconstruction, city and building modeling for constructing 3D virtual reality models because of their complementary nature. However, to gain from this complementarity, these data sources must be precisely registered. In this paper, we propose a new primitive based registration algorithm that takes 3D segments as features. The objective of the proposed algorithm is to register heterogeneous data. The heterogeneity is both in data type (image and LiDAR) and acquisition platform (terrestrial and aerial). Our algorithm starts by extracting 3D segments from LiDAR and image data with state of the art algorithms. Then it clusters the 3D segments of each data according to their directions. The obtained clusters are associated to find possible rotations, then 3D segments from associated clusters are matched in order to find the translation and scale factor minimizing a distance criteria between the two sets of 3D segments. Two optimizers (simulated annealing and RANSAC) are tested to minimize this distance criterion, first on synthetic data, then on real data. The experiments carried out demonstrate the robustness and speed of RANSAC compared to simulated annealing.

Indoor/Outdoor modeling of buildings is an important issue in the field of building life cycle management. It is seen as a joint process where the two aspects collaborate to take advantage of their semantic and geometric complementary. This global approach will allow a more complete, correct, precise and coherent reconstruction of the buildings . The first issue of such modeling is thus to precisely register this data. The lack of overlap between indoor and outdoor data is the most encountered obstacle, more so when both data sets are acquired separately and using different types of sensors. As an opening in the façade is the unique common entity that can be seen from inside and outside, it can help the registration of indoor and outdoor point clouds. So it must be automatically, accurately and efficiently extracted. In this paper,we start by proposing a very efficient algorithm to detect openings with great precision in both indoor and outdoor scans. Afterwards, we integrate them in a registration framework. As an opening is defined by a rectangular shape composed of four segments, two horizontal and two vertical, we can write our registration problem as a minimization of a global robust distance between two segment sets and propose a robust approach to minimize this distance using the RANSAC paradigm.

The registration of indoor and outdoor scans with a precision reaching the level of geometric noise represents a major challenge for Indoor/Outdoor building modeling. The basic idea of the contribution presented in this paper consists in extracting planar polygons from indoor and outdoor LiDAR scans, and then matching them. In order to cope with the very small overlap between indoor and outdoor scans of the same building, we propose to start by extracting points lying in the buildings’ interior from the outdoor scans as points where the laser ray crosses detected façades. Since, within a building environment, most of the objects are bounded by a planar surface, we propose a new registration algorithm that matches planar polygons by clustering polygons according to their normal direction, then by their offset in the normal direction. We use this clustering to find possible polygon correspondences (hypotheses) and estimate the optimal transformation for each hypothesis. Finally, a quality criteria is computed for each hypothesis in order to select the best one. To demonstrate the accuracy of our algorithm, we tested it on real data with a static indoor acquisition and a dynamic (Mobile Laser Scanning) outdoor acquisition.

This paper addresses the evaluation of algorithms reconstructing a watertight surface from a point cloud acquired on an open scene. The objective is to set a rigorous protocol measuring the quality of the reconstruction and to propose a quality metric that is informative with respect to the various qualities that such an algorithm should have, and in particular its capacity to interpolate and extrapolate accurately. Our approach aims at being more informative and rigorous than previous works on this topic. In addition, we use publicly available data and our implementation is open-source. We argue that a rigorous evaluation of surface reconstruction of open scenes needs to be performed on synthetic data where a perfect continuous ground truth surface is available, so we developed our own LiDAR simulator of which we give a description in the present paper.

Stereo dense matching is a fundamental task for 3D scene reconstruction. Recently, deep learning based methods have proven effective on some benchmark datasets, for example Middlebury and KITTI stereo. However, it is not easy to find a training dataset for aerial photogrammetry. Generating ground truth data for real scenes is a challenging task. In the photogrammetry community, many evaluation methods use digital surface models (DSM) to generate the ground truth disparity for the stereo pairs, but in this case interpolation may bring errors in the estimated disparity. In this paper, we publish a stereo dense matching dataset based on ISPRS Vaihingen dataset, and use it to evaluate some traditional and deep learning based methods. The evaluation shows that learning-based methods outperform traditional methods significantly when the fine tuning is done on a similar landscape. The benchmark also investigates the impact of the base to height ratio on the performance of the evaluated methods. The dataset can be found in https://github.com/whuwuteng/benchmark_ISPRS2021.

The availability of 3D Geospatial information is a key issue for many expanding sectors such as autonomous vehicles, business intelligence and urban planning. Its production is now possible thanks to the abundance of available data (Earth observation satellite constellations, insitu data, …) but manual interventions are still needed to guarantee a high level of quality, which prevents mass production. New artificial intelligence and big data technologies adapted to 3D imagery can help to remove these obstacles. The AI4GEO project aims at developing an automatic solution for producing 3D geospatial information and new added-value services. This paper will first introduce AI4GEO initiative, context and overall objectives. It will then present the current status of the project and in particular it will focus on the innovative platform put in place to handle big 3D datasets for analytics needs and it will present the first results of 3D semantic segmentations and associated perspectives.

Detecting planar structures in point clouds is a very central step of the point cloud processing pipeline as many Lidar scans, in particular in anthropic environments, present such planar structures. Many improvements have been proposed to RANSAC and the Hough transform, the two major types of plane detection methods. An important limitation however is that these methods detect planes running across the whole scene instead of more localized planar patches. Moreover, they do not exploit the sensor information that often comes with Lidar point cloud (sensor topology and optical center position in particular). In this paper we address both issues: we aim at detecting planar polygons that have a limited spatial extent, and we exploit sensor topology. The latter is used to enhance a RANSAC framework on two aspects: to make seed points selection more local and to define more compact sets of inliers through sensor space region growing.

Indoor mapping attracts more attention with the development of 2D and 3D camera and Lidar sensor. Lidar systems can provide a very high resolution and accurate point cloud. When aiming to reconstruct the static part of the scene, moving objects should be detected and removed which can prove challenging. This paper proposes a generic method to merge meshes produced from Lidar data that allows to tackle the issues of moving objects removal and static scene reconstruction at once. The method is adapted to a platform collecting point cloud from two Lidar sensors with different scan direction, which will result in different quality. Firstly, a mesh is efficiently produced from each sensor by exploiting its natural topology. Secondly, a visibility analysis is performed to handle occlusions (due to varying viewpoints) and remove moving objects. Then, a boolean optimization allows to select which triangles should be removed from each mesh. Finally, a stitching method is used to connect the selected mesh pieces. Our method is demonstrated on a Navvis M3 (2D laser ranger system) dataset and compared with Poisson and Delaunay based reconstruction methods.

We introduce a new method for the piecewise-planar approximation of 3D data, including point clouds and meshes. Our method is designed to operate on large datasets (e.g. millions of vertices) containing planar structures, which are very frequent in anthropic scenes. Our approach is also adaptive to the local geometric complexity of the input data. Our main contribution is the formulation of the piecewise-planar approximation problem as a non-convex optimization problem. In turn, this problem can be efficiently solved with a graph-structured working set approach. We compare our results with a state-of-the-art region-growing-based segmentation method and show a significant improvement both in terms of approximation error and computation efficiency.

We propose a new method for the reconstruction of simplicial complexes (combining points, edges and triangles) from 3D point clouds from Mobile Laser Scanning (MLS). Our main goal is to produce a reconstruction of a scene that is adapted to the local geometry of objects. Our method uses the inherent topology of the MLS sensor to define a spatial adjacency relationship between points. We then investigate each possible connexion between adjacent points and filter them by searching collinear structures in the scene, or structures perpendicular to the laser beams. Next, we create triangles for each triplet of self-connected edges. Last, we improve this method with a regularization based on the co-planarity of triangles and collinearity of remaining edges. We compare our results to a naive simplicial complexes reconstruction based on edge length.

The representation of 3D geometric and photometric information of the real world is one of the most challenging and extensively studied research topics in the photogrammetry and robotics communities. In this paper, we present a fully automatic framework for 3D high quality large scale urban texture mapping using oriented images and LiDAR scans acquired by a terrestrial Mobile Mapping System (MMS). First, the acquired points and images are sliced into temporal chunks ensuring a reasonable size and time consistency between geometry (points) and photometry (images). Then, a simple, fast and scalable 3D surface reconstruction relying on the sensor space topology is performed on each chunk after an isotropic sampling of the point cloud obtained from the raw LiDAR scans. Finally, the algorithm proposed in (Waechter et al., 2014) is adapted to texture the reconstructed surface with the images acquired simultaneously, ensuring a high quality texture with no seams and global color adjustment. We evaluate our full pipeline on a dataset of 17 km of acquisition in Rouen, France resulting in nearly 2 billion points and 40000 full HD images. We are able to reconstruct and texture the whole acquisition in less than 30 computing hours, the entire process being highly parallel as each chunk can be processed independently in a separate thread or computer.

Lidar technology has been widely used in both robotics and geomatics for environment perception and mapping. Moving object detection is important in both fields as it is a fundamental step for collision avoidance, static background extraction, moving pattern analysis, etc. A simple method involves checking directly the distance between nearest points from the compared datasets. However, large distances may be obtained when two datasets have different coverages. The use of occupancy grids is a popular approach to overcome this problem. There are two common theories employed to model occupancy and to interpret the measurements, Dempster- Shafer theory and probability. This paper presents a comparative study of these two theories for occupancy modelling with the aim of moving object detection from lidar point clouds. Occupancy is modelled using both approaches and their implementations are explained and compared in details. Two lidar datasets are tested to illustrate the moving object detection results.

Pedestrian traffic flow estimation is essential for public place design and construction planning. Traditional data collection by human investigation is tedious, inefficient and expensive. Panoramic laser scanners, e.g. Velodyne HDL-64E, which scan surroundings repetitively at a high frequency, have been increasingly used for 3D object tracking. In this paper, a simultaneous detection and tracking (SDAT) method is proposed for precise and automatic pedestrian trajectory recovery. First, the dynamic environment is detected using two different methods, Nearest-point and Max-distance. Then, all the points on moving objects are transferred into a space-time (x, y, t) coordinate system. The pedestrian detection and tracking amounts to assign the points belonging to pedestrians into continuous trajectories in space-time. We formulate the point assignment task as an energy function which incorporates the point evidence, trajectory number, pedestrian shape and motion. A low energy trajectory will well explain the point observations, and have plausible trajectory trend and length. The method inherently filters out points from other moving objects and false detections. The energy function is solved by a two-step optimization process: tracklet detection in a short temporal window; and global tracklet association through the whole time span. Results demonstrate that the proposed method can automatically recover the pedestrians trajectories with accurate positions and low false detections and mismatches.

This paper proposes an hybrid online calibration method for a laser scanner mounted on a mobile platform also equipped with an imaging system. The method relies on finding the calibration parameters that best align the acquired points cloud to the images. The quality of this intermodal alignment is measured by Mutual information between image luminance and points reflectance. The main advantage and motivation is ensuring pixel accurate alignment of images and point clouds acquired simultaneously, but it is also much more flexible than traditional laser calibration methods.

Laser scanners are used more and more in mobile mapping systems. They provide 3D point clouds that are used for object reconstruction and registration of the system. For both of those applications, uncertainty analysis of 3D points is of great interest but rarely investigated in the literature. In this paper we present a complete pipeline that takes into account all the sources of uncertainties and allows to compute a covariance matrix per 3D point. The sources of uncertainties are laser scanner, calibration of the scanner in relation to the vehicle and direct georeferencing system. We suppose that all the uncertainties follow the Gaussian law. The variances of the laser scanner measurements (two angles and one distance) are usually evaluated by the constructors. This is also the case for integrated direct georeferencing devices. Residuals of the calibration process were used to estimate the covariance matrix of the 6D transformation between scanner laser and the vehicle system. Knowing the variances of all sources of uncertainties, we applied uncertainty propagation technique to compute the variance-covariance matrix of every obtained 3D point. Such an uncertainty analysis enables to estimate the impact of different laser scanners and georeferencing devices on the quality of obtained 3D points. The obtained uncertainty values were illustrated using error ellipsoids on different datasets.

Techniques for 3D mapping are maturing. At the same time the need for 3D data is increasing. This has pushed national (and regional) mapping agencies (NMAs) to consider extending their traditional task of providing topographic data into the third dimension. To show how research results in 3D mapping obtained over the past twenty years have been adopted by practice, this paper presents the ongoing work on 3D mapping within seven NMAs, all member of the 3D Special Interest Group of European Spatial Data Research (EuroSDR). The paper shows that some NMAs are still in the initial (experimental) phase of 3D mapping, while others have already built solid databases to maintain 2.5D and 3D topographic data covering their whole country.

This paper proposes a pipeline to produce road orthophoto and DTM from Mobile Laser Scanning (MLS). For the ortho, modern laser scanners provide a reflectance information allowing for high quality grayscale images, at a much finer resolution than aerial photography can offer. For DTM, MLS offers a much higher accuracy and density than aerial products. This increased precision and resolution leverages new applications for both ortho and DEM. The first task is to filter ground vs non ground, then an interpolation is conducted to build image tiles from the filtered points. Finally, multiple layers are registered and blended to allow for seamless fusion. Our proposed approach achieves high quality products and scaling up is demonstrated.

Despite of the popularity of Delauney structure for mesh generation, octree based approaches remain an interesting solution for a first step surface reconstruction. In this paper, we propose a generic framework for a octree cell based mesh generation. Its input is a set of Lidar-based 3D measurements or other inputs which are formulated as a set of mass functions that characterize the level of confidence on the occupancy of each octree’s leaf. The output is a binary segmentation of the space between occupied and empty areas by taking into account the uncertainty of data. To this end, the problem is then reduced to a global energy optimization framework efficiently optimized with a min-cut approach. We use the approach for producing a large scale surface reconstruction algorithm by merging data from ubiquitous sources like airborne, terrestrial Lidar data, occupancy map and extra cues. Once the surface is computed, a solution is proposed for texturing the mesh.

Mobile Mapping Systems (MMS) are now commonly acquiring lidar scans of urban environments for an increasing number of applications such as 3D reconstruction and mapping, urban planning, urban furniture monitoring, practicability assessment for persons with reduced mobility (PRM)... MMS acquisitions are usually huge enough to incur a usability bottleneck for the increasing number of non-expert user that are not trained to process and visualize these huge datasets through specific softwares. A vast majority of their current need is for a simple 2D visualization that is both legible on screen and printable on a static 2D medium, while still conveying the understanding of the 3D scene and minimizing the disturbance of the lidar acquisition geometry (such as lidar shadows). The users that motivated this research are, by law, bound to precisely georeference underground networks for which they currently have schematics with no or poor absolute georeferencing. A solution that may fit their needs is thus a 2D visualization of the MMS dataset that they could easily interpret and on which they could accurately match features with their user datasets they would like to georeference. Our main contribution is two-fold. First, we propose a 3D point cloud stylization for 2D static visualization that leverages a Principal Component Analysis (PCA)-like local geometry analysis. By skipping the usual and error-prone estimation of a ground elevation, this rendering is thus robust to non-flat areas and has no hard-to-tune parameters such as height thresholds. Second, we implemented the corresponding rendering pipeline so that it can scale up to arbitrary large datasets by leveraging the Spark framework and its Resilient Distributed Dataset (RDD) and Dataframe abstractions.

This paper presents a full pipeline to extract mobile objects in images based on a simultaneous laser acquisition with a Velodyne scanner. The point cloud is first analysed to extract mobile objects in 3D. This is done using Dempster-Shafer theory and it results in weights telling for each points if it corresponds to a mobile object, a fixed object or if no decision can be made based on the data (unknown). These weights are projected in an image acquired simultaneously and used to segment the image between the mobile and the static part of the scene.

The 3D reconstruction of similar 3D objects detected in 2D faces a major issue when it comes to grouping the 2D detections into clusters to be used to reconstruct the individual 3D objects. Simple clustering heuristics fail as soon as similar objects are close. This paper formulates a framework to use the geometric quality of the reconstruction as a hint to do a proper clustering. We present a methodology to solve the resulting combinatorial optimization problem with some simplifications and approximations in order to make it tractable. The proposed method is applied to the reconstruction of 3D traffic signs from their 2D detections to demonstrate its capacity to solve ambiguities.

Historical photographs become widely used in geographical and environmental applications. Their enhancement involves converting them into georeferenced data, such as orthoimages or digital models. However no ground control points are available unlike in current image processing, and many problems such as image noise, landscape modifications, perspective distortion and unknown sensor calibration prevent automatic tie-point retrieval with current orthoimages. That is why photograph georeferencing remains a manual and time-consuming task. A novel method is presented in this paper to register photographs with current topographic database using line feature matching. Indeed, geometrical considerations only let avoid high radiometric difference issues when dealing with current orthoimages. Besides topographic database use lets selecting stable through time features, such as road network and historical buildings. A multi-scale approach allows very coarse georeferencing initialization, which can be set manually by a minimum number of ground control points per image set. At each scale an iterative processing improves the line matching and the registration model estimation at the same time. Finally, building integration makes registration more reliable for off-ground objects. Results are promising as georeferencing is much improved and its estimation converges in all test cases.

Reconstructing fine facade geometry from MMS lidar data remains a challenge: In addition to being inherently sparse, the point cloud provided by a single street point of view is necessarily incomplete. We propose a simple framework to estimate the facade surface with a deformable 2.5d grid. Computations are performed in a "sensor-oriented" coordinate system that maximizes consistency with the data. the algorithm allows to retrieve the facade geometry without priori knowledge. It can thus be automatically applied to a large amount of data in spite of the variability of encountered architectural forms. The 2.5d image structure of the output makes it compatible with storage and real-time constraints of immersive navigation.

Digital Elevation Models provide an accurate description of scenes allowing change detection by radiometric independence measurements. The height information gives access to the changes in the 3D shape of a scene occurring between two dates. In this paper, a procedure to extract meaningful changes is proposed based on analysing this shape variation. Instead of filtering a difference map, a multilevel analysis of the difference map based on the theory of persistence is developed to make such analysis independent from a choice of threshold values.

This article presents a generic and efficient method to register terrestrial mobile data with imperfect location on a geographic database with better overall accuracy but less details. The registration method proposed in this paper is based on a semi-rigid point to plane ICP ("Iterative Closest Point"). The main applications of such registration is to improve existing geographic databases, particularly in terms of accuracy, level of detail and diversity of represented objects. Other applications include fine geometric modelling and fine façade texturing, object extraction such as trees, poles, road signs marks, facilities, vehicles, etc. The geopositionning system of mobile mapping systems is affected by GPS masks that are only partially corrected by an Inertial Navigation System (INS) which can cause an important drift. As this drift varies non-linearly, but slowly in time, it will be modelled by a translation defined as a piecewise linear function of time which variation over time will be minimized (rigidity term). For each iteration of the ICP, the drift is estimated in order to minimise the distance between laser points and planar model primitives (data attachment term). The method has been tested on real data (a scan of the city of Paris of 3.6 million laser points registered on a 3D model of approximately 71,400 triangles).

Thanks to the development of Mobile mapping systems (MMS), street object recognition, classification, modelling and related studies have become hot topics recently. There has been increasing interest in detecting changes between mobile laser scanning (MLS) point clouds in complex urban areas. A method based on the consistency between the occupancies of space computed from different datasets is proposed. First occupancy of scan rays (empty, occupied, unknown) are defined while considering the accuracy of measurement and registration. Then the occupancy of scan rays are fused using the Weighted Dempster–Shafer theory (WDST). Finally, the consistency between different datasets is obtained by comparing the occupancy at points from one dataset with the fused occupancy of neighbouring rays from the other dataset. Change detection results are compared with a conventional point to triangle (PTT) distance method. Changes at point level are detected fully automatically. The proposed approach allows to detect changes at large scales in urban scenes with fine detail and more importantly, distinguish real changes from occlusions.

This paper presents a method to efficiently compute the visibility of a 3D model seen from a series of calibrated georeferenced images. This is required to make methods aimed at enriching 3D models (finer reconstruction, texture mapping) based on such images scalable. The method consists in rasterizing the scene on the GPU from the point of views of the images with a known mapping between colors and scene elements. This pixel based visibility information is then made higher level by defining a set of geometric rules to select which images should be used when focusing on a given scene element. The visibility relation is finally used to build a visibility graph allowing for efficient sequential processing of the whole scene

This paper proposes a method to estimate the local sharpness of an optical system through the wavelet-based analysis of a large set of images it acquired. Assuming a space-invariant distribution of image features, such as in the aerial photography context, the proposed approach produces a sharpness map of the imaging device over 16 × 16 pixels blocks that enables, for instance, the detection of optical defects and the qualification of the mosaicking of multiple sensor images into a larger composite image. The proposed analysis is based on accumulating of the edge maps corresponding to the first levels of the Haar Transform of each image of the dataset, following the intuition that statistically, each pixel will see the same image structures. We propose a calibration method to transform these accumulated edge maps into a sharpness map by approximating the local PSF (Point Spread Function) with a Gaussian blur.

This papers presents a multi-scale method that computes robust geometric features on lidar point clouds in order to retrieve the optimal neighborhood size for each point. Three dimensionality features are calculated on spherical neighborhoods at various radius sizes. Based on combinations of the eigenvalues of the local structure tensor, they describe the shape of the neighborhood, indicating whether the local geometry is more linear (1D), planar (2D) or volumetric (3D). Two radius-selection criteria have been tested and compared for finding automatically the optimal neighborhood radius for each point. Besides, such procedure allows a dimensionality labelling, giving significant hints for classification and segmentation purposes. The method is successfully applied to 3D point clouds from airborne, terrestrial, and mobile mapping systems since no a priori knowledge on the distribution of the 3D points is required. Extracted dimensionality features and labellings are then favorably compared to those computed from constant size neighborhoods.

3D reconstruction of trees is of great interest in large-scale 3D city modelling. Laser scanners provide geometrically accurate 3D point clouds that are very useful for object recognition in complex urban scenes. Trees often cause important occlusions on building façades. Their recognition can lead to occlusion maps that are useful for many façade oriented applications such as visual based localisation and automatic image tagging. This paper proposes a pipeline to detect trees in point clouds acquired in dense urban areas with only laser informations (x,y, z coordinates and intensity). It is based on local geometric descriptors computed on each laser point using a determined neighbourhood. These descriptors describe the local shape of objects around every 3D laser point. A projection of these values on a 2D horizontal accumulation space followed by a combination of morphological filters provides individual tree clusters. The pipeline is evaluated and the results are presented on a set of one million laser points using a man made ground truth.

This article describes a pipeline developed to automatically detect and correct motion blur due to the airplane motion in aerial images provided by a digital camera system with channel-dependent exposure times. Blurred images show anisotropy in their Fourier Transform coefficients that can be detected and estimated to recover the characteristics of the motion blur. To disambiguate the anisotropy produced by a motion blur from the possible spectral anisotropy produced by some periodic patterns present in a sharp image, we consider the phase difference of the Fourier Transform of two channel shot with different exposure times (i.e. with different blur extensions). This is possible because of the deep correlation between the three visible channels ensures phase coherence of the Fourier Transform coefficients in sharp images. In this context, considering the phase difference constitutes both a good detector and estimator of the motion blur parameters. In order to improve on this estimation, the phase difference is performed on local windows in the image where the channels are more correlated. The main lobe of the phase difference, where the phase difference between two channels is close to zero actually imitates an ellipse which axis ratio discriminates blur and which orientation and minor axis give respectively the orientation and the blur kernel extension of the long exposure-time channels. However, this approach is not robust to the presence in the phase difference of minor lobes due to phase sign inversions in the Fourier transform of the motion blur. They are removed by considering the polar representation of the phase difference. Based on the blur detection step, blur correction is eventually performed using two different approaches depending on the blur extension size: using either a simple frequency-based fusion for small blur or a semi blind iterative method for larger blur. The higher computing costs of the latter method make it only suitable for large motion blur, when the former method is not applicable.

A reliable and accurate facade database would be a major asset in applications such as localization of autonomous vehicles, registration and fine building modeling. Mobile mapping devices now provide the data required to create such a database, but efficient methods should be designed in order to tackle the enormous amount of data collected by such means (a million point per second for hours of acquisition). Another important limitation is the presence of numerous objects in urban scenes of many different types. This paper proposes a method that overcomes these two issues: – The facade detection algorithm is streamed: the data is processed in the order it was acquired. More precisely, the input data is split into overlapping blocks which are analysed in turn to extract facade parts. Close overlapping parts are then merged in order to recover the full facade rectangle. – The geometry of the neighborhood of each point is analysed to define a probability that the point belongs to a vertical planar patch. This probability is then injected in a RANdom SAmple Consensus (RANSAC) algorithm both in the sampling step and in the hypothesis validation, in order to favour the most reliable candidates. This ensures much more robustness against outliers during the facade detection. This way, the main vertical rectangles are detected without any prior knowledge about the data. The only assumptions are that the facades are roughly planar and vertical. The method has been successfully tested on a large dataset in Paris. The facades are detected despite the presence of trees occluding large areas of some facades. The robustness and accuracy of the detected facade rectangles makes them useful for localization applications and for registration of other scans of the same city or of entire city models.