Search

International audience; Traditional Convolutional Neural Networks (CNN) for semantic segmentation of images use 2D convolution operations. While the spatial inductive bias of 2D convolutions allow CNNs to build hierarchical feature representations, they require that the whole feature maps are kept in memory until the end of the inference. This is not ideal for memory and latency-critical applications such as real-time on-board satellite image segmentation. In this paper, we propose a new neural network architecture for semantic segmentation, "Scan-nerNet", based on a Recurrent 1D Convolutional architecture. Our network performs a segmentation of the input image lineby-line, and thus reduces the memory footprint and output latency. These characteristics make it ideal for on-the-fly segmentation of images on-board satellites equipped with push broom sensors such as Landsat 8, or satellites with limited compute capabilities, such as Cubesats. We perform cloud segmentation experiments on embedded hardware and show that our method offers a good compromise between accuracy, memory usage and latency.

International audience; We present an algorithm for detecting planar primitives from unorganized 3D point clouds. Departing from an initial configuration, the algorithm refines both the continuous plane parameters and the discrete assignment of input points to them by seeking high fidelity, high simplicity and high completeness. Our key contribution relies upon the design of an exploration mechanism guided by a multiobjective energy function. The transitions within the large solution space are handled by five geometric operators that create, remove and modify primitives. We demonstrate the potential of our method on a variety of scenes, from organic shapes to man-made objects, and sensors, from multiview stereo to laser. We show its efficacy with respect to existing primitive fitting approaches and illustrate its applicative interest in compact mesh reconstruction, when combined with a plane assembly method.

International audience; Automatic road graph extraction from aerial and satellite images is a long-standing challenge. Existing algorithms are either based on pixel-level segmentation followed by vectorization, or on iterative graph construction using next move prediction. Both of these strategies suffer from severe drawbacks, in particular high computing resources and incomplete outputs. By contrast, we propose a method that directly infers the final road graph in a single pass. The key idea consists in combining a Fully Convolutional Network in charge of locating points of interest such as intersections, dead ends and turns, and a Graph Neural Network which predicts links between these points. Such a strategy is more efficient than iterative methods and allows us to streamline the training process by removing the need for generation of starting locations while keeping the training end-to-end. We evaluate our method against existing works on the popular RoadTracer dataset and achieve competitive results. We also benchmark the speed of our method and show that it outperforms existing approaches. Our method opens the possibility of in-flight processing on embedded devices for applications such as real-time road network monitoring and alerts for disaster response.

International audience; We present a global registration algorithm for multi-modal geometric data, typically 3D point clouds and meshes. Existing feature-based methods and recent deep learning based approaches typically rely upon point-to-point matching strategies that often fail to deliver accurate results from defect-laden data. In contrast, we reason at the scale of planar shapes whose detection from input data offers robustness on a range of defects, from noise to outliers through heterogeneous sampling. The detected planar shapes are projected into an accumulation space from which a rotational alignment is operated. A second step then refines the result with a local continuous optimization which also estimates the scale. We demonstrate the robustness and efficacy of our algorithm on challenging real-world data. In particular, we show that our algorithm competes well against state-of-the-art methods, especially on piece-wise planar objects and scenes.

International audience

The automatic generation of 3D urban models from geospa-tial data is now a standard procedure. However, practitioners still have to visually assess, at city-scale, the correctness of these models and detect inevitable reconstruction errors. Such a process relies on experts, and is highly time-consuming (2 h/km 2 /expert). In this work, we propose an approach for automatically evaluating the quality of 3D building models. Potential errors are compiled in a hierarchical, versatile and parameter-izable taxonomy. This allows for the first time to disentangle fidelity and modeling errors, whatever the level of details of the modeled buildings. The quality of models is predicted using the geometric properties of buildings and, when available, image and depth data. A baseline of handcrafted, yet generic features, is fed to a Random Forest classifier. Both multi-class and multi-label cases are considered. Due to the interdependence between classes of errors, we have the ability to retrieve all errors at the same time while predicting erroneous buildings. We tested our framework on an urban area with more than 1,000 building models. We can satisfactorily detect, on average 96% of the most frequent errors.

This habilitation thesis proposes a series of contributions in the field of geometric modeling from physical measurements. These contributions present concepts and algorithms for object extraction, surface reconstruction, and scene modeling for urban environments. The methodology behind these contributions relies on stochastic geometry and graphical models. These probabilistic models are suited to analyzing the diversity and the complexity of urban objects and to exploring large and highly non-convex solution spaces at city-scales. The physical measurements, data structures, and concepts involved in our work lie between computer vision, geometry processing, and photogrammetry. A general summary also provides a vision of the many remaining challenges in the field.

International audience; We present a novel approach for the decimation of trian-gle surface meshes. Our algorithm takes as input a triangle surface mesh and a set of planar proxies detected in a pre-processing analysis step, and structured via an adjacency graph. It then performs greedy mesh decimation through a series of edge collapse, designed to approximate the local mesh geometry as well as the geometry and structure of proxies. Such structure-preserving approach is well suited to planar abstraction, i.e., extreme decimation approxi-mating well the planar parts while filtering out the others. Our experiments on a variety of inputs illustrate the po-tential of our approach in terms of improved accuracy and preservation of structure.

International audience; We present a method for planar shape detection and regularization from raw point sets. The geometric modeling and processing of man-made environments from measurement data often relies upon robust detection of planar primitive shapes. In addition, the detection and reinforcement of regularities between planar parts is a means to increase resilience to missing or defect-laden data as well as to reduce the complexity of models and algorithms down the modeling pipeline. The main novelty behind our method is to perform detection and regularization in tandem. We first sample a sparse set of seeds uniformly on the input point set, then perform in parallel shape detection through region growing, interleaved with regularization through detection and reinforcement of regular relationships (coplanar, parallel and orthogonal). In addition to addressing the end goal of regularization, such reinforcement also improves data fitting and provides guidance for clustering small parts into larger planar parts. We evaluate our approach against a wide range of inputs and under four criteria: geometric fidelity, coverage, regularity and running times. Our approach compares well with available implementations such as the efficient RANSAC-based approach proposed by Schnabel and co-authors in 2007.

This habilitation thesis proposes a series of contributions in the field of geometric modeling from physical measurements. These contributions present concepts and algorithms for object extraction, surface reconstruction, and scene modeling for urban environments. The methodology behind these contributions relies on stochastic geometry and graphical models. These probabilistic models are suited to analyzing the diversity and the complexity of urban objects and to exploring large and highly non-convex solution spaces at city-scales. The physical measurements, data structures, and concepts involved in our work lie between computer vision, geometry processing, and photogrammetry. A general summary also provides a vision of the many remaining challenges in the field.

International audience; We present a method for automatic reconstruction of permanent structures of indoor scenes, such as walls, floors and ceilings, from raw point clouds acquired by laser scanners. Our approach employs graph-cut to solve an inside/outside labeling of a space decomposition. To allow for an accurate reconstruction the space decomposition is aligned with permanent structures. A Hough Transform is applied for extracting the wall directions while allowing a flexible reconstruction of scenes. The graph-cut formulation takes into account data consistency through an inside/outside prediction for the cells of the space decomposition by stochastic ray casting, while favoring low geometric complexity of the model. Our experiments produces watertight reconstructed models of multi-level buildings and complex scenes.