Yvain Quéau

Nathan Buskulic, Jalal Fadili, Yvain Quéau

Neural networks have become a prominent approach to solve inverse problems in recent years. While a plethora of such methods was developed to solve inverse problems empirically, we are still lacking clear theoretical guarantees for these methods. On the other hand, many works proved convergence to optimal solutions of neural networks in a more general setting using overparametrization as a way to control the Neural Tangent Kernel. In this work we investigate how to bridge these two worlds and we provide deterministic convergence and recovery guarantees for the class of unsupervised feedforward multilayer neural networks trained to solve inverse problems. We also derive overparametrization bounds under which a two-layers Deep Inverse Prior network with smooth activation function will benefit from our guarantees.

Nathan Buskulic, Yvain Quéau, Jalal Fadili

Neural networks have become a prominent approach to solve inverse problems in recent years. Amongst the different existing methods, the Deep Image/Inverse Priors (DIPs) technique is an unsupervised approach that optimizes a highly overparametrized neural network to transform a random input into an object whose image under the forward model matches the observation. However, the level of overparametrization necessary for such methods remains an open problem. In this work, we aim to investigate this question for a two-layers neural network with a smooth activation function. We provide overparametrization bounds under which such network trained via continuous-time gradient descent will converge exponentially fast with high probability which allows to derive recovery prediction bounds. This work is thus a first step towards a theoretical understanding of overparametrized DIP networks, and more broadly it participates to the theoretical understanding of neural networks in inverse problem settings.

Robin Bruneau, Baptiste Brument, Yvain Quéau et al.

Achieving high-fidelity 3D surface reconstruction while preserving fine details remains challenging, especially in the presence of materials with complex reflectance properties and without a dense-view setup. In this paper, we introduce a versatile framework that incorporates multi-view normal and optionally reflectance maps into radiance-based surface reconstruction. Our approach employs a pixel-wise joint re-parametrization of reflectance and surface normals, representing them as a vector of radiances under simulated, varying illumination. This formulation enables seamless incorporation into standard surface reconstruction pipelines, such as traditional multi-view stereo (MVS) frameworks or modern neural volume rendering (NVR) ones. Combined with the latter, our approach achieves state-of-the-art performance on multi-view photometric stereo (MVPS) benchmark datasets, including DiLiGenT-MV, LUCES-MV and Skoltech3D. In particular, our method excels in reconstructing fine-grained details and handling challenging visibility conditions. The present paper is an extended version of the earlier conference paper by Brument et al. (in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024), featuring an accelerated and more robust algorithm as well as a broader empirical evaluation. The code and data relative to this article is available at https://github.com/RobinBruneau/RNb-NeuS2.

Clément Hardy, Yvain Quéau, David Tschumperlé

The photometric stereo (PS) problem consists in reconstructing the 3D-surface of an object, thanks to a set of photographs taken under different lighting directions. In this paper, we propose a multi-scale architecture for PS which, combined with a new dataset, yields state-of-the-art results. Our proposed architecture is flexible: it permits to consider a variable number of images as well as variable image size without loss of performance. In addition, we define a set of constraints to allow the generation of a relevant synthetic dataset to train convolutional neural networks for the PS problem. Our proposed dataset is much larger than pre-existing ones, and contains many objects with challenging materials having anisotropic reflectance (e.g. metals, glass). We show on publicly available benchmarks that the combination of both these contributions drastically improves the accuracy of the estimated normal field, in comparison with previous state-of-the-art methods.

Nathan Buskulic, Jalal Fadil, Yvain Quéau

Solving inverse problems with neural networks benefits from very few theoretical guarantees when it comes to the recovery guarantees. We provide in this work convergence and recovery guarantees for self-supervised neural networks applied to inverse problems, such as Deep Image/Inverse Prior, and trained with inertia featuring both viscous and geometric Hessian-driven dampings. We study both the continuous-time case, i.e., the trajectory of a dynamical system, and the discrete case leading to an inertial algorithm with an adaptive step-size. We show in the continuous-time case that the network can be trained with an optimal accelerated exponential convergence rate compared to the rate obtained with gradient flow. We also show that training a network with our inertial algorithm enjoys similar recovery guarantees though with a less sharp linear convergence rate.

Nathan Buskulic, Jalal Fadili, Yvain Quéau

Advanced machine learning methods, and more prominently neural networks, have become standard to solve inverse problems over the last years. However, the theoretical recovery guarantees of such methods are still scarce and difficult to achieve. Only recently did unsupervised methods such as Deep Image Prior (DIP) get equipped with convergence and recovery guarantees for generic loss functions when trained through gradient flow with an appropriate initialization. In this paper, we extend these results by proving that these guarantees hold true when using gradient descent with an appropriately chosen step-size/learning rate. We also show that the discretization only affects the overparametrization bound for a two-layer DIP network by a constant and thus that the different guarantees found for the gradient flow will hold for gradient descent.

Matéo Ducastel, David Tschumperlé, Yvain Quéau

Recent state-of-the-art algorithms in photometric stereo rely on neural networks and operate either through prior learning or inverse rendering optimization. Here, we revisit the problem of calibrated photometric stereo by leveraging recent advances in 3D inverse rendering using the Gaussian Splatting formalism. This allows us to parameterize the 3D scene to be reconstructed and optimize it in a more interpretable manner. Our approach incorporates a simplified model for light representation and demonstrates the potential of the Gaussian Splatting rendering engine for the photometric stereo problem.

Mohammed Brahimi, Yvain Quéau, Bjoern Haefner et al.

Uncalibrated photometric stereo aims at estimating the 3D-shape of a surface, given a set of images captured from the same viewing angle, but under unknown, varying illumination. While the theoretical foundations of this inverse problem under directional lighting are well-established, there is a lack of mathematical evidence for the uniqueness of a solution under general lighting. On the other hand, stable and accurate heuristical solutions of uncalibrated photometric stereo under such general lighting have recently been proposed. The quality of the results demonstrated therein tends to indicate that the problem may actually be well-posed, but this still has to be established. The present paper addresses this theoretical issue, considering first-order spherical harmonics approximation of general lighting. Two important theoretical results are established. First, the orthographic integrability constraint ensures uniqueness of a solution up to a global concave-convex ambiguity, which had already been conjectured, yet not proven. Second, the perspective integrability constraint makes the problem well-posed, which generalizes a previous result limited to directional lighting. Eventually, a closed-form expression for the unique least-squares solution of the problem under perspective projection is provided, allowing numerical simulations on synthetic data to empirically validate our findings.

Bjoern Haefner, Zhenzhang Ye, Maolin Gao et al.

Photometric stereo (PS) techniques nowadays remain constrained to an ideal laboratory setup where modeling and calibration of lighting is amenable. To eliminate such restrictions, we propose an efficient principled variational approach to uncalibrated PS under general illumination. To this end, the Lambertian reflectance model is approximated through a spherical harmonic expansion, which preserves the spatial invariance of the lighting. The joint recovery of shape, reflectance and illumination is then formulated as a single variational problem. There the shape estimation is carried out directly in terms of the underlying perspective depth map, thus implicitly ensuring integrability and bypassing the need for a subsequent normal integration. To tackle the resulting nonconvex problem numerically, we undertake a two-phase procedure to initialize a balloon-like perspective depth map, followed by a "lagged" block coordinate descent scheme. The experiments validate efficiency and robustness of this approach. Across a variety of evaluations, we are able to reduce the mean angular error consistently by a factor of 2-3 compared to the state-of-the-art.

Bjoern Haefner, Songyou Peng, Alok Verma et al.

This study explores the use of photometric techniques (shape-from-shading and uncalibrated photometric stereo) for upsampling the low-resolution depth map from an RGB-D sensor to the higher resolution of the companion RGB image. A single-shot variational approach is first put forward, which is effective as long as the target's reflectance is piecewise-constant. It is then shown that this dependency upon a specific reflectance model can be relaxed by focusing on a specific class of objects (e.g., faces), and delegate reflectance estimation to a deep neural network. A multi-shot strategy based on randomly varying lighting conditions is eventually discussed. It requires no training or prior on the reflectance, yet this comes at the price of a dedicated acquisition setup. Both quantitative and qualitative evaluations illustrate the effectiveness of the proposed methods on synthetic and real-world scenarios.

Georg Radow, Laurent Hoeltgen, Yvain Quéau et al.

Estimating shape and appearance of a three dimensional object from a given set of images is a classic research topic that is still actively pursued. Among the various techniques available, PS is distinguished by the assumption that the underlying input images are taken from the same point of view but under different lighting conditions. The most common techniques provide the shape information in terms of surface normals. In this work, we instead propose to minimise a much more natural objective function, namely the reprojection error in terms of depth. Minimising the resulting non-trivial variational model for PS allows to recover the depth of the photographed scene directly. As a solving strategy, we follow an approach based on a recently published optimisation scheme for non-convex and non-smooth cost functions. The main contributions of our paper are of theoretical nature. A technical novelty in our framework is the usage of matrix differential calculus. We supplement our approach by a detailed convergence analysis of the resulting optimisation algorithm and discuss possibilities to ease the computational complexity. At hand of an experimental evaluation we discuss important properties of the method. Overall, our strategy achieves more accurate results than competing approaches. The experiments also highlights some practical aspects of the underlying optimisation algorithm that may be of interest in a more general context.

Yvain Quéau, Jean Mélou, Fabien Castan et al.

A numerical solution to shape-from-shading under natural illumination is presented. It builds upon an augmented Lagrangian approach for solving a generic PDE-based shape-from-shading model which handles directional or spherical harmonic lighting, orthographic or perspective projection, and greylevel or multi-channel images. Real-world applications to shading-aware depth map denoising, refinement and completion are presented.

Jean Mélou, Yvain Quéau, Jean-Denis Durou et al.

We tackle the problem of reflectance estimation from a set of multi-view images, assuming known geometry. The approach we put forward turns the input images into reflectance maps, through a robust variational method. The variational model comprises an image-driven fidelity term and a term which enforces consistency of the reflectance estimates with respect to each view. If illumination is fixed across the views, then reflectance estimation remains under-constrained: a regularization term, which ensures piecewise-smoothness of the reflectance, is thus used. Reflectance is parameterized in the image domain, rather than on the surface, which makes the numerical solution much easier, by resorting to an alternating majorization-minimization approach. Experiments on both synthetic and real datasets are carried out to validate the proposed strategy.

Yvain Quéau, Jean-Denis Durou, Jean-François Aujol

The need for an efficient method of integration of a dense normal field is inspired by several computer vision tasks, such as shape-from-shading, photometric stereo, deflectometry, etc. Inspired by edge-preserving methods from image processing, we study in this paper several variational approaches for normal integration, with a focus on non-rectangular domains, free boundary and depth discontinuities. We first introduce a new discretization for quadratic integration, which is designed to ensure both fast recovery and the ability to handle non-rectangular domains with a free boundary. Yet, with this solver, discontinuous surfaces can be handled only if the scene is first segmented into pieces without discontinuity. Hence, we then discuss several discontinuity-preserving strategies. Those inspired, respectively, by the Mumford-Shah segmentation method and by anisotropic diffusion, are shown to be the most effective for recovering discontinuities.

Yvain Quéau, Jean-Denis Durou, Jean-François Aujol

The need for efficient normal integration methods is driven by several computer vision tasks such as shape-from-shading, photometric stereo, deflectometry, etc. In the first part of this survey, we select the most important properties that one may expect from a normal integration method, based on a thorough study of two pioneering works by Horn and Brooks [28] and by Frankot and Chellappa [19]. Apart from accuracy, an integration method should at least be fast and robust to a noisy normal field. In addition, it should be able to handle several types of boundary condition, including the case of a free boundary, and a reconstruction domain of any shape i.e., which is not necessarily rectangular. It is also much appreciated that a minimum number of parameters have to be tuned, or even no parameter at all. Finally, it should preserve the depth discontinuities. In the second part of this survey, we review most of the existing methods in view of this analysis, and conclude that none of them satisfies all of the required properties. This work is complemented by a companion paper entitled Variational Methods for Normal Integration, in which we focus on the problem of normal integration in the presence of depth discontinuities, a problem which occurs as soon as there are occlusions.

Songyou Peng, Bjoern Haefner, Yvain Quéau et al.

A novel depth super-resolution approach for RGB-D sensors is presented. It disambiguates depth super-resolution through high-resolution photometric clues and, symmetrically, it disambiguates uncalibrated photometric stereo through low-resolution depth cues. To this end, an RGB-D sequence is acquired from the same viewing angle, while illuminating the scene from various uncalibrated directions. This sequence is handled by a variational framework which fits high-resolution shape and reflectance, as well as lighting, to both the low-resolution depth measurements and the high-resolution RGB ones. The key novelty consists in a new PDE-based photometric stereo regularizer which implicitly ensures surface regularity. This allows to carry out depth super-resolution in a purely data-driven manner, without the need for any ad-hoc prior or material calibration. Real-world experiments are carried out using an out-of-the-box RGB-D sensor and a hand-held LED light source.

Yvain Quéau, Bastien Durix, Tao Wu et al.

We conduct a thorough study of photometric stereo under nearby point light source illumination, from modeling to numerical solution, through calibration. In the classical formulation of photometric stereo, the luminous fluxes are assumed to be directional, which is very difficult to achieve in practice. Rather, we use light-emitting diodes (LEDs) to illuminate the scene to reconstruct. Such point light sources are very convenient to use, yet they yield a more complex photometric stereo model which is arduous to solve. We first derive in a physically sound manner this model, and show how to calibrate its parameters. Then, we discuss two state-of-the-art numerical solutions. The first one alternatingly estimates the albedo and the normals, and then integrates the normals into a depth map. It is shown empirically to be independent from the initialization, but convergence of this sequential approach is not established. The second one directly recovers the depth, by formulating photometric stereo as a system of PDEs which are partially linearized using image ratios. Although the sequential approach is avoided, initialization matters a lot and convergence is not established either. Therefore, we introduce a provably convergent alternating reweighted least-squares scheme for solving the original system of PDEs, without resorting to image ratios for linearization. Finally, we extend this study to the case of RGB images.

Yvain Quéau, Jean Mélou, Jean-Denis Durou et al.

We introduce a variational method for multi-view shape-from-shading under natural illumination. The key idea is to couple PDE-based solutions for single-image based shape-from-shading problems across multiple images and multiple color channels by means of a variational formulation. Rather than alternatingly solving the individual SFS problems and optimizing the consistency across images and channels which is known to lead to suboptimal results, we propose an efficient solution of the coupled problem by means of an ADMM algorithm. In numerous experiments on both simulated and real imagery, we demonstrate that the proposed fusion of multiple-view reconstruction and shape-from-shading provides highly accurate dense reconstructions without the need to compute dense correspondences. With the proposed variational integration across multiple views shape-from-shading techniques become applicable to challenging real-world reconstruction problems, giving rise to highly detailed geometry even in areas of smooth brightness variation and lacking texture.

Martin Bähr, Michael Breuß, Yvain Quéau et al.

The integration of surface normals for the purpose of computing the shape of a surface in 3D space is a classic problem in computer vision. However, even nowadays it is still a challenging task to devise a method that combines the flexibility to work on non-trivial computational domains with high accuracy, robustness and computational efficiency. By uniting a classic approach for surface normal integration with modern computational techniques we construct a solver that fulfils these requirements. Building upon the Poisson integration model we propose to use an iterative Krylov subspace solver as a core step in tackling the task. While such a method can be very efficient, it may only show its full potential when combined with a suitable numerical preconditioning and a problem-specific initialisation. We perform a thorough numerical study in order to identify an appropriate preconditioner for our purpose. To address the issue of a suitable initialisation we propose to compute this initial state via a recently developed fast marching integrator. Detailed numerical experiments illuminate the benefits of this novel combination. In addition, we show on real-world photometric stereo datasets that the developed numerical framework is flexible enough to tackle modern computer vision applications.