Visual localization is the problem of estimating the position and orientation from which a given image (or a sequence of images) is taken in a known scene. It is an important part of a wide range of computer vision and robotics applications, from self-driving cars to augmented/virtual reality systems. Visual localization techniques should work reliably and robustly under a wide range of conditions, including seasonal, weather, illumination and man-made changes. Recent benchmarking efforts model this by providing images under different conditions, and the community has made rapid progress on these datasets since their inception. However, they are limited to a few geographical regions and often recorded with a single device. We propose a new benchmark for visual localization in outdoor scenes, using crowd-sourced data to cover a wide range of geographical regions and camera devices with a focus on the failure cases of current algorithms. Experiments with state-of-the-art localization approaches show that our dataset is very challenging, with all evaluated methods failing on its hardest parts. As part of the dataset release, we provide the tooling used to generate it, enabling efficient and effective 2D correspondence annotation to obtain reference poses.

Lifelong place recognition is an essential and challenging task in computer vision with vast applications in robust localization and efficient large-scale 3D reconstruction. Progress is currently hindered by a lack of large, diverse, publicly available datasets. We contribute with Mapillary Street-Level Sequences (SLS), a large dataset for urban and suburban place recognition from image sequences. It contains more than 1.6 million images curated from the Mapillary collaborative mapping platform. The dataset is orders of magnitude larger than current data sources, and is designed to reflect the diversities of true lifelong learning. It features images from 30 major cities across six continents, hundreds of distinct cameras, and substantially different viewpoints and capture times, spanning all seasons over a nine year period. All images are geo-located with GPS and compass, and feature high-level attributes such as road type. We propose a set of benchmark tasks designed to push state-of-the-art performance and provide baseline studies. We show that current state-of-the-art methods still have a long way to go, and that the lack of diversity in existing datasets have prevented generalization to new environments. The dataset and benchmarks are available for academic research.

Single image calibration is the problem of predicting the camera parameters from one image. This problem is of importance when dealing with images collected in uncontrolled conditions by non-calibrated cameras, such as crowd-sourced applications. In this work we propose a method to predict extrinsic (tilt and roll) and intrinsic (focal length and radial distortion) parameters from a single image. We propose a parameterization for radial distortion that is better suited for learning than directly predicting the distortion parameters. Moreover, predicting additional heterogeneous variables exacerbates the problem of loss balancing. We propose a new loss function based on point projections to avoid having to balance heterogeneous loss terms. Our method is, to our knowledge, the first to jointly estimate the tilt, roll, focal length, and radial distortion parameters from a single image. We thoroughly analyze the performance of the proposed method and the impact of the improvements and compare with previous approaches for single image radial distortion correction.

In this paper we propose an algorithm for stereoscopic image inpainting, given the inpainting mask in both images. We also assume that depth map is known in one of the images of the stereo pair, taken as reference. This image is clustered in homogeneous color regions using a mean-shift procedure. In each clustered region, depths are fitted by planes and then extended into the mask. Then we inpaint the visible parts of each extended region using a modified exemplar-based inpainting algorithm. Finally, we extend the algorithm to stereoscopic image inpainting. We display some experiments showing the performance of the proposed algorithm.

In this paper the problem of optical flow and occlusion mask estimation is aborded. To that end, we consider a multi-label representation of the optical flow and we define an energy that models the problem. The convexification of the energy and its minimization with an iterative algorithm are studied. Our algorithm is implemented in GPU, since each pixel can be processed in parallel. From our experiments, the relation between the quality of the results obtained and computing time seems to be very promising.

In this paper we propose an algorithm for in painting of stereo images. The issue is to reconstruct the holes in a pair of stereo image as if they were the projection of a 3D scene. Hence, the reconstruction of the missing information has to produce a consistent visual perception of depth. Thus, first step of the algorithm consists in the computation and in painting of disparity maps in the given holes. The second step of the algorithm is to fill-in missing regions using the complete disparity maps in a way that avoids the creation of 3D artifacts. We present some experiments on several pairs of stereo images.

International audience

This paper addresses the problem of image-based surface reconstruction. The main contribution is the computation of the exact derivative of the reprojection error functional. This allows its rigorous minimization via gradient descent surface evolution. The main difficulty has been to correctly take into account the visibility changes that occur when the surface moves. A geometric and analytical study of these changes is presented and used for the computation of derivative. Our analysis shows the strong influence that the movement of the contour generators has on the reprojection error. As a consequence, during the proper minimization of the reprojection error, the contour generators of the surface are automatically moved to their correct location in the images. Therefore, current methods adding additional silhouettes or apparent contour constraints to ensure this alignment can now be understood and justified by a single criterion: the reprojection error.

This paper addresses the problem of reconstructing the geometry and color of a Lambertian scene, given some fully calibrated images acquired with wide baselines. In order to completely model the input data, we propose to represent the scene as a set of colored depth maps, one per input image. We formulate the problem as a Bayesian MAP problem which leads to an energy minimization method. Hidden visibility variables are used to deal with occlusion, reflections and outliers. The main contributions of this work are: a prior for the visibility variables that treats the geometric occlusions; and a prior for the multiple depth maps model that smoothes and merges the depth maps while enabling discontinuities. Real world examples showing the efficiency and limitations of the approach are presented.