Depth estimation is of critical interest for scene understanding and accurate 3D reconstruction. Most recent approaches in depth estimation with deep learning exploit geometrical structures of standard sharp images to predict corresponding depth maps. However, cameras can also produce images with defocus blur depending on the depth of the objects and camera settings. Hence, these features may represent an important hint for learning to predict depth. In this paper, we propose a full system for single-image depth prediction in the wild using depth-fromdefocus and neural networks. We carry out thorough experiments to test deep convolutional networks on real and simulated defocused images using a realistic model of blur variation with respect to depth. We also investigate the influence of blur on depth prediction observing model uncertainty with a Bayesian neural network approach. From these studies, we show that out-of-focus blur greatly improves the depth-prediction network performances. Furthermore, we transfer the ability learned on a synthetic, indoor dataset to real, indoor and outdoor images. For this purpose, we present a new dataset containing real all-focus and defocused images from a Digital Single-Lens Reflex (DSLR) camera, paired with ground truth depth maps obtained with an active 3D sensor for indoor scenes. The proposed approach is successfully validated on both this new dataset and standard ones as NYUv2 or Depth-in-the-Wild. Code and new datasets are available at https:// github.com/ marcelampc/ d3net depth estimation.

Depth estimation from a single monocular image has reached great performances thanks to recent works based on deep networks. However, as various choices of losses, architectures and experimental conditions are proposed in the literature, it is difficult to establish their respective influence on the performances. In this paper we propose an in-depth study of various losses and experimental conditions for depth regression, on NYUv2 dataset. From this study we propose a new network for depth estimation combining an encoder-decoder architecture with an adversarial loss. This network reaches top ones state of the art on NUYv2 dataset while being simpler to train in a single phase.

We present a new dataset, dedicated to the development of simultaneous localization and mapping methods for underwater vehicles navigating close to the seabed. The data sequences composing this dataset are recorded in three different environments: a harbor at a depth of a few meters, a first archaeological site at a depth of 270 meters and a second site at a depth of 380 meters. The data acquisition is performed using Remotely Operated Vehicles equipped with a monocular monochromatic camera, a low-cost inertial measurement unit, a pressure sensor and a computing unit, all embedded in a single enclosure. The sensors' measurements are recorded synchronously on the computing unit and seventeen sequences have been created from all the acquired data. These sequences are made available in the form of ROS bags and as raw data. For each sequence, a trajectory has also been computed offline using a Structure-from-Motion library in order to allow the comparison with real-time localization methods. With the release of this dataset, we wish to provide data difficult to acquire and to encourage the development of vision-based localization methods dedicated to the underwater environment. The dataset can be downloaded from: http://www.lirmm.fr/aqualoc/

In the context of underwater robotics, the visual degradation induced by the medium properties make difficult the exclusive use of cameras for localization purpose. Hence, many underwater localization methods are based on expensive navigation sensors associated with acoustic positioning. On the other hand, pure visual localization methods have shown great potential in underwater localization but the challenging conditions, such as the presence of turbidity and dynamism, remain complex to tackle. In this paper, we propose a new visual odometry method designed to be robust to these visual perturbations. The proposed algorithm has been assessed on both simulated and real underwater datasets and outperforms state-of-the-art terrestrial visual SLAM methods under many of the most challenging conditions. The main application of this work is the localization of Remotely Operated Vehicles used for underwater archaeological missions, but the developed system can be used in any other applications as long as visual information is available.

In this paper we propose a new monocular depth estimation algorithm based on local estimation of defocus blur, approach referred to as Depth from Defocus (DFD). Using a limited set of calibration images, we directly learn image covariance which indeed encode both scene and blur (i.e. depth) information. Depth is then estimated from a single image patch using a maximum likelihood criterion defined using the learned covariance. This method is applied here within a new active DFD method using a dense textured projection and a chromatic lens for image acquisition. The projector adds texture for low textured objects -which is usually a limitation of DFD -and the chromatic aberration increases the estimated depth range with respect to conventional DFD. We provide here quantitative evaluations of the depth estimation performance of our method on simulated and real data of fronto-parallel untextured scenes. The proposed method is then qualitatively experimentally evaluated on 3D-printed benchmark.

In this paper we propose a new concept for a compact 3D sensor dedicated to industrial inspection, combining chromatic Depth From Defocus (DFD) and structured illumination. Depth is estimated from a single image using local estimation of the defocus blur. As industrial objects usually show poor texture information, which is crucial for DFD, we rely on structured illumination. In contrast with state of the art approaches for active DFD, which project sparse patterns on the scene, our method exploits a dense textured pattern and provides dense depth maps of the scene. Besides, to overcome depth ambiguity and dead zone of DFD with a classical camera, we use an unconventional lens with chromatic aberration, providing spectrally varying defocus blur in the camera color channels. We provide comparisons of depth estimation performance for several projected patterns at various scales based on simulation and real experiments. The proposed method is then qualitatively evaluated on a real industrial object. Finally we discuss the perspectives of this work especially in terms of co-design of an 3D active sensor using DFD.

Aerial or satellite imagery is a great source for land surface analysis, which might yield land use maps or elevation models. In this investigation, we present a neural network framework for learning semantics and local height together. We show how this joint multi-task learning benefits to each task on the large dataset of the 2018 Data Fusion Contest. Moreover, our framework also yields an uncertainty map which allows assessing the prediction of the model. Code is available at https://github.com/marcelampc/mtl_aerial_images .

We propose to add an optical component in front of a conventional camera to improve depth estimation performance of depth from defocus (DFD), an approach based on the relation between defocus blur and depth. The add-on overcomes ambiguity and the dead zone, which are the fundamental limitations of DFD with a conventional camera, by adding an optical aberration to the whole system that makes the blur unambiguous and measurable for each depth. We look into two optical components: the first one adds astigmatism and the other one chromatic aberration. In both cases, we present the principle of the add-on and experimental validations on real prototypes.