We present the design of a productionized end-to-end stereo depth sensing system that does pre-processing, online stereo rectification, and stereo depth estimation with a fallback to monocular depth estimation when rectification is unreliable. The output of our depth sensing system is then used in a novel view generation pipeline to create 3D computational photography effects using point-of-view images captured by smart glasses. All these steps are executed on-device on the stringent compute budget of a mobile phone, and because we expect the users can use a wide range of smartphones, our design needs to be general and cannot be dependent on a particular hardware or ML accelerator such as a smartphone GPU. Although each of these steps is well studied, a description of a practical system is still lacking. For such a system, all these steps need to work in tandem with one another and fallback gracefully on failures within the system or less than ideal input data. We show how we handle unforeseen changes to calibration, e.g., due to heat, robustly support depth estimation in the wild, and still abide by the memory and latency constraints required for a smooth user experience. We show that our trained models are fast, and run in less than 1s on a six-year-old Samsung Galaxy S8 phone's CPU. Our models generalize well to unseen data and achieve good results on Middlebury and in-the-wild images captured from the smart glasses.

Inferring human-scene contact (HSC) is the first step toward understanding how humans interact with their surroundings. While detecting 2D human-object interaction (HOI) and reconstructing 3D human pose and shape (HPS) have enjoyed significant progress, reasoning about 3D human-scene contact from a single image is still challenging. Existing HSC detection methods consider only a few types of predefined contact, often reduce the body and scene to a small number of primitives, and even overlook image evidence. To predict human-scene contact from a single image, we address the limitations above from both data and algorithmic perspectives. We capture a new dataset called RICH for “Real scenes, Interaction, Contact and Humans.” RICH contains multiview outdoor/indoor video sequences at 4K resolution, ground-truth 3D human bodies captured using markerless motion capture, 3D body scans, and high resolution 3D scene scans. A key feature of RICH is that it also contains accurate vertex-level contact labels on the body. Using RICH, we train a network that predicts dense body-scene contacts from a single RGB image. Our key insight is that regions in contact are always occluded so the network needs the ability to explore the whole image for evidence. We use a transformer to learn such non-local relationships and propose a new Body-Scene contact TRansfOrmer (BSTRO). Very few methods explore 3D contact; those that do focus on the feet only, detect foot contact as a post-processing step, or infer contact from body pose without looking at the scene. To our knowledge, BSTRO is the first method to directly estimate 3D body-scene contact from a single image. We demonstrate that BSTRO significantly outperforms the prior art. Our code and dataset are available for research purposes at: https://rich.is.tue.mpg.de.

Semi-Global Matching (SGM) is a widely-used efficient stereo matching technique. It works well for textured scenes, but fails on untextured slanted surfaces due to its fronto-parallel smoothness assumption. To remedy this problem, we propose a simple extension, termed SGM-P, to utilize precomputed surface orientation priors. Such priors favor different surface slants in different 2D image regions or 3D scene regions and can be derived in various ways. In this paper we evaluate plane orientation priors derived from stereo matching at a coarser resolution and show that such priors can yield significant performance gains for difficult weakly-textured scenes. We also explore surface normal priors derived from Manhattan-world assumptions, and we analyze the potential performance gains using oracle priors derived from ground-truth data. SGM-P only adds a minor computational overhead to SGM and is an attractive alternative to more complex methods employing higher-order smoothness terms.

Untextured scenes with complex occlusions still present challenges to modern stereo algorithms. We consider the pathological case of Mondrian Stereo-scenes consisting solely of solid-colored planar regions, inspired by paintings by Piet Mondrian. We analyze assumptions that allow disambiguating such scenes and present a novel stereo algorithm employing symbolic reasoning about matched edge segments. We demonstrate compelling stereo matching results on synthetic scenes and discuss how our insights could be utilized in robust real-world stereo algorithms for untextured environments.

Local fusion of disparity maps allows fast parallel 3D modeling of large scenes that do not fit into main memory. While existing methods assume a constant disparity uncertainty, disparity errors typically vary spatially from tenths of pixels to several pixels. In this paper we propose a method that employs a set of Gaussians for different disparity classes, instead of a single error model with only one variance. The set of Gaussians is learned from the difference between generated disparity maps and ground-truth disparities. Pixels are assigned particular disparity classes based on a Total Variation (TV) feature measuring the local oscillation behavior of the 2D disparity map. This feature captures uncertainty caused for instance by lack of texture or fron to-parallel bias of the stereo method. Experimental results on several datasets in varying configurations demonstrate that our method yields improved performance both qualitatively and quantitatively.

We present a stereo algorithm designed for speed and efficiency that uses local slanted plane sweeps to propose disparity hypotheses for a semi-global matching algorithm. Our local plane hypotheses are derived from initial sparse feature correspondences followed by an iterative clustering step. Local plane sweeps are then performed around each slanted plane to produce out-of-plane parallax and matching-cost estimates. A final global optimization stage, implemented using semi-global matching, assigns each pixel to one of the local plane hypotheses. By only exploring a small fraction of the whole disparity space volume, our technique achieves significant speedups over previous algorithms and achieves state-of-the-art accuracy on high-resolution stereo pairs of up to 19 megapixels.

To produce images that are suitable for display, tone-mapping is widely used in digital cameras to map linear color measurements into narrow gamuts with limited dynamic range. This introduces non-linear distortion that must be undone, through a radiometric calibration process, before computer vision systems can analyze such photographs radiometrically. This paper considers the inherent uncertainty of undoing the effects of tone-mapping. We observe that this uncertainty varies substantially across color space, making some pixels more reliable than others. We introduce a model for this uncertainty and a method for fitting it to a given camera or imaging pipeline. Once fit, the model provides for each pixel in a tone-mapped digital photograph a probability distribution over linear scene colors that could have induced it. We demonstrate how these distributions can be useful for visual inference by incorporating them into estimation algorithms for a representative set of vision tasks.

We propose a novel image-based rendering algorithm for handling complex scenes that may include reflective surfaces. Our key contribution lies in treating the problem in the gradient domain. We use a standard technique to estimate scene depth, but assign depths to image gradients rather than pixels. A novel view is obtained by rendering the horizontal and vertical gradients, from which the final result is reconstructed through Poisson integration using an approximate solution as a data term. Our algorithm is able to handle general scenes including reflections and similar effects without explicitly separating the scene into reflective and transmissive parts, as required by previous work. Our prototype renderer is fully implemented on the GPU and runs in real time on commodity hardware.

We present a system for image-based modeling and rendering of real-world scenes containing reflective and glossy surfaces. Previous approaches to image-based rendering assume that the scene can be approximated by 3D proxies that enable view interpolation using traditional back-to-front or z-buffer compositing. In this work, we show how these can be generalized to multiple layers that are combined in an additive fashion to model the reflection and transmission of light that occurs at specular surfaces such as glass and glossy materials. To simplify the analysis and rendering stages, we model the world using piecewise-planar layers combined using both additive and opaque mixing of light. We also introduce novel techniques for estimating multiple depths in the scene and separating the reflection and transmission components into different layers. We then use our system to model and render a variety of real-world scenes with reflections.

This paper presents a method for joint stereo matching and object segmentation. In our approach a 3D scene is represented as a collection of visually distinct and spatially coherent objects. Each object is characterized by three different aspects: a color model, a 3D plane that approximates the object's disparity distribution, and a novel 3D connectivity property. Inspired by Markov Random Field models of image segmentation, we employ object-level color models as a soft constraint, which can aid depth estimation in powerful ways. In particular, our method is able to recover the depth of regions that are fully occluded in one input view, which to our knowledge is new for stereo matching. Our model is formulated as an energy function that is optimized via fusion moves. We show high-quality disparity and object segmentation results on challenging image pairs as well as standard benchmarks. We believe our work not only demonstrates a novel synergy between the areas of image segmentation and stereo matching, but may also inspire new work in the domain of automatic and interactive object-level scene manipulation.

The quantitative evaluation of optical flow algorithms by Barron et al. (1994) led to significant advances in performance. The challenges for optical flow algorithms today go beyond the datasets and evaluation methods proposed in that paper. Instead, they center on problems associated with complex natural scenes, including nonrigid motion, real sensor noise, and motion discontinuities. We propose a new set of benchmarks and evaluation methods for the next generation of optical flow algorithms. To that end, we contribute four types of data to test different aspects of optical flow algorithms: (1) sequences with nonrigid motion where the ground-truth flow is determined by tracking hidden fluorescent texture, (2) realistic synthetic sequences, (3) high frame-rate video used to study interpolation error, and (4) modified stereo sequences of static scenes. In addition to the average angular error used by Barron et al., we compute the absolute flow endpoint error, measures for frame interpolation error, improved statistics, and results at motion discontinuities and in textureless regions. In October 2007, we published the performance of several well-known methods on a preliminary version of our data to establish the current state of the art. We also made the data freely available on the web at http://vision.middlebury.edu/flow/ . Subsequently a number of researchers have uploaded their results to our website and published papers using the data. A significant improvement in performance has already been achieved. In this paper we analyze the results obtained to date and draw a large number of conclusions from them.

We present a new method for the robust detection and matching of multiple planes in pairs of images. Such planes can serve as stable landmarks for vision-based urban navigation. Our approach starts from SIFT matches and generates multiple local homography hypotheses using the recent J-linkage technique by Toldo and Fusiello, a robust randomized multi-model estimation algorithm. These hypotheses are then globally merged, spatially analyzed, robustly fitted, and checked for stability. When tested on more than 30,000 image pairs taken from panoramic views of a college campus, our method yields no false positives and recovers 72% of the matchable building walls identified by a human, despite significant occlusions and viewpoint changes.

Images harvested from the Web are proving to be useful for many visual tasks, including recognition, geo-location, and three-dimensional reconstruction. These images are captured under a variety of lighting conditions by consumer-level digital cameras, and these cameras have color processing pipelines that are diverse, complex, and scenedependent. As a result, the color information contained in these images is difficult to exploit. In this paper, we analyze the factors that contribute to the color output of a typical camera, and we explore the use of parametric models for relating these output colors to meaningful scenes properties. We evaluate these models using a database of registered images captured with varying camera models, camera settings, and lighting conditions. The database is available online at

Stereo correspondence methods rely on matching costs for computing the similarity of image locations. We evaluate the insensitivity of different costs for passive binocular stereo methods with respect to radiometric variations of the input images. We consider both pixel-based and window-based variants like the absolute difference, the sampling-insensitive absolute difference, and normalized cross correlation, as well as their zero-mean versions. We also consider filters like LoG, mean, and bilateral background subtraction (BilSub) and nonparametric measures like Rank, SoftRank, Census, and Ordinal. Finally, hierarchical mutual information (HMI) is considered as pixelwise cost. Using stereo data sets with ground-truth disparities taken under controlled changes of exposure and lighting, we evaluate the costs with a local, a semiglobal, and a global stereo method. We measure the performance of all costs in the presence of simulated and real radiometric differences, including exposure differences, vignetting, varying lighting, and noise. Overall, the ranking of methods across all data sets and experiments appears to be consistent. Among the best costs are BilSub, which performs consistently very well for low radiometric differences; HMI, which is slightly better as pixelwise matching cost in some cases and for strong image noise; and Census, which showed the best and most robust overall performance.

Among the most exciting advances in early vision has been the development of efficient energy minimization algorithms for pixel-labeling tasks such as depth or texture computation. It has been known for decades that such problems can be elegantly expressed as Markov random fields, yet the resulting energy minimization problems have been widely viewed as intractable. Recently, algorithms such as graph cuts and loopy belief propagation (LBP) have proven to be very powerful: for example, such methods form the basis for almost all the top-performing stereo methods. However, the tradeoffs among different energy minimization algorithms are still not well understood. In this paper we describe a set of energy minimization benchmarks and use them to compare the solution quality and running time of several common energy minimization algorithms. We investigate three promising recent methods graph cuts, LBP, and tree-reweighted message passing in addition to the well-known older iterated conditional modes (ICM) algorithm. Our benchmark problems are drawn from published energy functions used for stereo, image stitching, interactive segmentation, and denoising. We also provide a general-purpose software interface that allows vision researchers to easily switch between optimization methods. Benchmarks, code, images, and results are available at http://vision.middlebury.edu/MRF/.

Stereo correspondence methods rely on matching costs for computing the similarity of image locations. In this paper we evaluate the insensitivity of different matching costs with respect to radiometric variations of the input images. We consider both pixel-based and window-based variants and measure their performance in the presence of global intensity changes (e.g., due to gain and exposure differences), local intensity changes (e.g., due to vignetting, non-Lambertian surfaces, and varying lighting), and noise. Using existing stereo datasets with ground-truth disparities as well as six new datasets taken under controlled changes of exposure and lighting, we evaluate the different costs with a local, a semi-global, and a global stereo method.

State-of-the-art stereo vision algorithms utilize color changes as important cues for object boundaries. Most methods impose heuristic restrictions or priors on disparities, for example by modulating local smoothness costs with intensity gradients. In this paper we seek to replace such heuristics with explicit probabilistic models of disparities and intensities learned from real images. We have constructed a large number of stereo datasets with ground-truth disparities, and we use a subset of these datasets to learn the parameters of conditional random fields (CRFs). We present experimental results illustrating the potential of our approach for automatically learning the parameters of models with richer structure than standard hand-tuned MRF models.

The quantitative evaluation of optical flow algorithms by Barron et al. led to significant advances in the performance of optical flow methods. The challenges for optical flow today go beyond the datasets and evaluation methods proposed in that paper and center on problems associated with nonrigid motion, real sensor noise, complex natural scenes, and motion discontinuities. Our goal is to establish a new set of benchmarks and evaluation methods for the next generation of optical flow algorithms. To that end, we contribute four types of data to test different aspects of optical flow algorithms: sequences with nonrigid motion where the ground-truth flow is determined by tracking hidden fluorescent texture; realistic synthetic sequences; high frame-rate video used to study interpolation error; and modified stereo sequences of static scenes. In addition to the average angular error used in Barron et al., we compute the absolute flow endpoint error, measures for frame interpolation error, improved statistics, and flow accuracy at motion boundaries and in textureless regions. We evaluate the performance of several well-known methods on this data to establish the current state of the art. Our database is freely available on the web together with scripts for scoring and publication of the results at http://vision.middlebury.edu/flow/.

This paper presents a quantitative comparison of several multi-view stereo reconstruction algorithms. Until now, the lack of suitable calibrated multi-view image datasets with known ground truth (3D shape models) has prevented such direct comparisons. In this paper, we first survey multi-view stereo algorithms and compare them qualitatively using a taxonomy that differentiates their key properties. We then describe our process for acquiring and calibrating multiview image datasets with high-accuracy ground truth and introduce our evaluation methodology. Finally, we present the results of our quantitative comparison of state-of-the-art multi-view stereo reconstruction algorithms on six benchmark datasets. The datasets, evaluation details, and instructions for submitting new models are available online at http://vision.middlebury.edu/mview.

This paper presents a new method for navigation and localization of a mobile robot equipped with an omnidirectional camera. We represent the environment using a collection of one-dimensional panoramic images formed by averaging the center scanlines of a cylindrical view. Such 1D images can be stored and processed with few resources, allowing a fairly dense sampling of the environment. Image matching proceeds in real time using dynamic programming on scale-invariant features extracted from each circular view. By analyzing the shape of the matching curve, the relative orientation of pairs of views can be recovered and utilized for navigation. When navigating, the robot continually matches its current view against stored reference views taken from known locations, and determines its location and heading from the properties of the matching results. Experiments show that our method is robust to occlusion, repeating patterns, and lighting variations

A central issue in stereo algorithm design is the choice of matching cost. Many algorithms simply use squared or absolute intensity differences based on integer disparity steps. In this paper, we address potential problems with such approaches. We begin with a careful analysis of the properties of the continuous disparity space image (DSI) and propose several new matching cost variants based on symmetrically matching interpolated image signals. Using stereo images with ground truth, we empirically evaluate the performance of the different cost variants and show that proper sampling can yield improved matching performance.

Progress in stereo algorithm performance is quickly outpacing the ability of existing stereo data sets to discriminate among the best-performing algorithms, motivating the need for more challenging scenes with accurate ground truth information. This paper describes a method for acquiring high-complexity stereo image pairs with pixel-accurate correspondence information using structured light. Unlike traditional range-sensing approaches, our method does not require the calibration of the light sources and yields registered disparity maps between all pairs of cameras and illumination projectors. We present new stereo data sets acquired with our method and demonstrate their suitability for stereo algorithm evaluation. Our results are available at http://www.middlebury.edu/stereo/.

We propose a new system for vision-based mobile robot navigation in an unmodeled environment. Simple, unobtrusive artificial landmarks are used as navigation and localization aids. The landmark patterns are designed so that they can be reliably detected in real-time in images taken with the robot's camera over a wide range of viewing configurations. The code for the recognition algorithm is available in the web site.

We propose a new method for view synthesis from real images using stereo vision. The method does not explicitly model scene geometry, and enables fast and exact generation of synthetic views. We also reevaluate the requirements on stereo algorithms for the application of view synthesis and discuss ways of dealing with partially occluded regions of unknown depth and with completely occluded regions of unknown texture. Our experiments demonstrate that it is possible to efficiently synthesize realistic new views even from inaccurate and incomplete depth information.

One of the central problems in stereo matching (and other image registration tasks) is the selection of optimal window sizes for comparing image regions. This paper addresses this problem with some novel algorithms based on iteratively diffusing support at different disparity hypotheses, and locally controlling the amount of diffusion based on the current quality of the disparity estimate. It also develops a novel Bayesian estimation technique which significantly outperforms techniques based on area-based matching (SSD) and regular diffusion. We provide experimental results on both synthetic and real stereo image pairs.

We present a simple yet powerful method to perform point-to-point matching between two images. The method uses an evidence measure, whose value for a given displacement reflects both the similarity between two locations and the confidence in a correct match. The measure is based on the gradient fields of the images, and can be computed quickly and in parallel. Accumulating the evidence measure for different displacements allows (1) stable computation, of correspondences without smoothing across motion boundaries, and (2) detection of dominant motions. The method works well both on highly textured images and on images containing regions of uniform intensities, and can be used for a variety of applications, including stereo, motion, and object tracking.