When deploying deep neural networks on robots or other physical systems, the learned model should reliably quantify predictive uncertainty. A reliable uncertainty allows downstream modules to reason about the safety of its actions. In this work, we address metrics for uncertainty quantification. Specifically, we focus on regression tasks, and investigate Area Under Sparsification Error (AUSE), Calibration Error (CE), Spearman’s Rank Correlation, and Negative Log-Likelihood (NLL). Using multiple datasets, we look into how those metrics behave under four typical types of uncertainty, their stability regarding the size of the test set, and reveal their strengths and weaknesses. Our results indicate that Calibration Error is the most stable and interpretable metric, but AUSE and NLL also have their respective use cases. We discourage the usage of Spearman’s Rank Correlation for evaluating uncertainties and recommend replacing it with AUSE. • We explore evaluation metrics for uncertainty quantification. • We create toy datasets that highlight different sources of uncertainty. • Using our toy datasets, we compare and contrast metrics for uncertainty. • We evaluate: AUSE, Spearman Correlation, Calibration Error, and NLL. • Results: AUSE, NLL, Calibration error are good metrics with different strengths.

We tackle the task of scene flow estimation from point clouds. Given a source and a target point cloud, the objective is to estimate a translation from each point in the source point cloud to the target, resulting in a 3D motion vector field. Previous dominant scene flow estimation methods require complicated coarse-to-fine or recurrent architectures as a multi-stage refinement. In contrast, we propose a significantly simpler single-scale one-shot global matching to address the problem. Our key finding is that reliable feature similarity between point pairs is essential and sufficient to estimate accurate scene flow. We thus propose to decompose the feature extraction step via a hybrid local-global-cross transformer architecture which is crucial to accurate and robust feature representations. Extensive experiments show that the proposed Global Matching Scene Flow (GMSF) sets a new state-of-the-art on multiple scene flow estimation benchmarks. On FlyingThings3D, with the presence of occlusion points, GMSF reduces the outlier percentage from the previous best performance of 27.4% to 5.6%. On KITTI Scene Flow, without any fine-tuning, our proposed method shows state-of-the-art performance. On the Waymo-Open dataset, the proposed method outperforms previous methods by a large margin. The code is available at https://github.com/ZhangYushan3/GMSF.

We investigate spline-based continuous-time pose trajectory estimation using non-linear explicit motion priors. Current regularization priors either linearize the orientation, rely on the implicit regularization obtained from the used spline basis function, or use sampling based regularization schemes. The latter is a special case of a Riemann sum approximation, and we demonstrate when and why this can fail, and propose a way to avoid these issues. In addition we provide a number of novel practically useful theoretical contributions, including requirements on knot spacing for orientation splines, new basis functions for constant velocity extrapolation, and a generalization of the popular P-Spline penalty to orientation. We analyze the properties of the proposed approach using synthetic data. We validate our system using the standard task of visual-inertial calibration, and apply it to stereo visual odometry where we demonstrate real-time performance on KITTI.

Probabilistic methods for point set registration have interesting theoretical properties, such as linear complexity in the number of used points, and they easily generalize to joint registration of multiple point sets. In this work, we improve their recognition performance to match state of the art. This is done by incorporating learned features, by adding a von Mises-Fisher feature model in each mixture component, and by using learned attention weights. We learn these jointly using a registration loss learning strategy (RLL) that directly uses the registration error as a loss, by back-propagating through the registration iterations. This is possible as the probabilistic registration is fully differentiable, and the result is a learning framework that is truly end-to-end. We perform extensive experiments on the 3DMatch and Kitti datasets. The experiments demonstrate that our approach benefits significantly from the integration of the learned features and our learning strategy, outperforming the state-of-the-art on Kitti. Code is available at https://github.com/felja633/RLLReg.

Probabilistic methods for point set registration have demonstrated competitive results in recent years. These techniques estimate a probability distribution model of the point clouds. While such a representation has shown promise, it is highly sensitive to variations in the density of 3D points. This fundamental problem is primarily caused by changes in the sensor location across point sets. We revisit the foundations of the probabilistic registration paradigm. Contrary to previous works, we model the underlying structure of the scene as a latent probability distribution, and thereby induce invariance to point set density changes. Both the probabilistic model of the scene and the registration parameters are inferred by minimizing the Kullback-Leibler divergence in an Expectation Maximization based framework. Our density-adaptive registration successfully handles severe density variations commonly encountered in terrestrial Lidar applications. We perform extensive experiments on several challenging real-world Lidar datasets. The results demonstrate that our approach outperforms state-of-the-art probabilistic methods for multi-view registration, without the need of re-sampling.

In this paper we derive and test a probability-based weighting that can balance residuals of different types in spline fitting. In contrast to previous formulations, the proposed spline error weighting scheme also incorporates a prediction of the approximation error of the spline fit. We demonstrate the effectiveness of the prediction in a synthetic experiment, and apply it to visual-inertial fusion on rolling shutter cameras. This results in a method that can estimate 3D structure with metric scale on generic first-person videos. We also propose a quality measure for spline fitting, that can be used to automatically select the knot spacing. Experiments verify that the obtained trajectory quality corresponds well with the requested quality. Finally, by linearly scaling the weights, we show that the proposed spline error weighting minimizes the estimation errors on real sequences, in terms of scale and end-point errors.

The ability to direct visual attention is a fundamental skill for seeing robots. Attention comes in two flavours: the gaze direction (overt attention) and attention to a specific part of the current field of view (covert attention), of which the latter is the focus of the present study. Specifically, we study the effects of attentional masking within pre-trained deep neural networks for the purpose of handling ambiguous scenes containing multiple objects. We investigate several variants of attentional masking on partially pre-trained deep neural networks and evaluate the effects on classification performance and sensitivity to attention mask errors in multi-object scenes. We find that a combined scheme consisting of multi-level masking and blending provides the best trade-off between classification accuracy and insensitivity to masking errors. This proposed approach is denoted multilayer continuous-valued convolutional feature masking (MC-CFM). For reasonably accurate masks it can suppress the influence of distracting objects and reach comparable classification performance to unmasked recognition in cases without distractors.

We propose a technique for joint calibration of a wide-angle rolling shutter camera (e.g. a GoPro) and an externally mounted gyroscope. The calibrated parameters are time scaling and offset, relative pose between gyroscope and camera, and gyroscope bias. The parameters are found using non-linear least squares minimisation using the symmetric transfer error as cost function. The primary contribution is methods for robust initialisation of the relative pose and time offset, which are essential for convergence. We also introduce a robust error norm to handle outliers. This results in a technique that works with general video content and does not require any specific setup or calibration patterns. We apply our method to stabilisation of videos recorded by a rolling shutter camera, with a rigidly attached gyroscope. After recording, the gyroscope and camera are jointly calibrated using the recorded video itself. The recorded video can then be stabilised using the calibrated parameters. We evaluate the technique on video sequences with varying difficulty and motion frequency content. The experiments demonstrate that our method can be used to produce high quality stabilised videos even under difficult conditions, and that the proposed initialisation is shown to end up within the basin of attraction. We also show that a residual based on the symmetric transfer error is more accurate than residuals based on the recently proposed epipolar plane normal coplanarity constraint, and that the use of robust errors is a critical component to obtain an accurate calibration.

We propose an algorithm that can capture sharp, low-noise images in low-light conditions on a hand-held smartphone. We make use of the recent ability to acquire bursts of high resolution images on high-end models such as the iPhone5s. Frames are aligned, or stacked, using rolling shutter correction, based on motion estimated from the built-in gyro sensors and image feature tracking. After stacking, the images may be combined, using e.g. averaging to produce a sharp, low-noise photo. We have tested the algorithm on a variety of different scenes, using several different smartphones. We compare our method to denoising, direct stacking, as well as a global-shutter based stacking, with favourable results.

This paper deals with fast and accurate visualization of pushbroom image data from airborne and spaceborne platforms. A pushbroom sensor acquires images in a line-scanning fashion, and this results in scattered input data that need to be resampled onto a uniform grid for geometrically correct visualization. To this end, we model the anisotropic spatial dependence structure caused by the acquisition process. Several methods for scattered data interpolation are then adapted to handle the induced anisotropic metric and compared for the pushbroom image rectification problem. A trick that exploits the semiordered line structure of pushbroom data to improve the computational complexity several orders of magnitude is also presented.

This paper describes a method for generation of dense stereo ground-truth using a consumer depth sensor such as the Microsoft Kinect. Such ground-truth allows adaptation of stereo algorithms to a specific setting. The method uses a novel residual weighting based on error propagation from image plane measurements to 3D. We use this ground-truth in wide-angle stereo learning by automatically tuning a novel extension of the best-first-propagation (BFP) dense correspondence algorithm. We extend BFP by adding a coarse-to-fine scheme, and a structure measure that limits propagation along linear structures and flat areas. The tuned correspondence algorithm is evaluated in terms of accuracy, robustness, and ability to generalise. Both the tuning cost function, and the evaluation are designed to balance the accuracy-robustness trade-off inherent in patch-based methods such as BFP.

This paper introduces a bundle adjustment (BA) method that obtains accurate structure and motion from rolling shutter (RS) video sequences: RSBA. When a classical BA algorithm processes a rolling shutter video, the resultant camera trajectory is brittle, and complete failures are not uncommon. We exploit the temporal continuity of the camera motion to define residuals of image point trajectories with respect to the camera trajectory. We compare the camera trajectories from RSBA to those from classical BA, and from classical BA on rectified videos. The comparisons are done on real video sequences from an iPhone 4, with ground truth obtained from a global shutter camera, rigidly mounted to the iPhone 4. Compared to classical BA, the rolling shutter model requires just six extra parameters. It also degrades the sparsity of the system Jacobian slightly, but as we demonstrate, the increase in computation time is moderate. Decisive advantages are that RSBA succeeds in cases where competing methods diverge, and consistently produces more accurate results.

Structured light range sensors, such as the Microsoft Kinect, have recently become popular as perception devices for computer vision and robotic systems. These sensors use CMOS imaging chips with electronic rolling shutters (ERS). When using such a sensor on a moving platform, both the image, and the depth map, will exhibit geometric distortions. We introduce an algorithm that can suppress such distortions, by rectifying the 3D point clouds from the range sensor. This is done by first estimating the time continuous 3D camera trajectory, and then transforming the 3D points to where they would have been, if the camera had been stationary. To ensure that image and range data are synchronous, the camera trajectory is computed from KLT tracks on the structured-light frames, after suppressing the structured-light pattern. We evaluate our rectification, by measuring angles between the visible sides of a cube, before and after rectification. We also measure how much better the 3D point clouds can be aligned after rectification. The obtained improvement is also related to the actual rotational velocity, measured using a MEMS gyroscope.

This paper presents a method for rectifying video sequences from rolling shutter (RS) cameras. In contrast to previous RS rectification attempts we model distortions as being caused by the 3D motion of the camera. The camera motion is parametrised as a continuous curve, with knots at the last row of each frame. Curve parameters are solved for using non-linear least squares over inter-frame correspondences obtained from a KLT tracker. We have generated synthetic RS sequences with associated ground-truth to allow controlled evaluation. Using these sequences, we demonstrate that our algorithm improves over to two previously published methods. The RS dataset is available on the web to allow comparison with other methods.

Recent years have seen advances in the estimation of full 6 degree-of-freedom object pose from a single 2D image. These advances have often been presented as a result of, or together with, a new local image descriptor. This paper examines how the performance for such a system varies with choice of local descriptor. This is done by comparing the performance of a full 6 degree-of-freedom pose estimation system for fourteen types of local descriptors. The evaluation is done on a database with photos of complex objects with simple and complex backgrounds and varying lighting conditions. From the experiments we can conclude that duplet features, that use pairs of interest points, improve pose estimation accuracy, and that affine covariant features do not work well in current pose estimation frameworks. The data sets and their ground truth is available on the Web to allow future comparison with novel algorithms.

This paper studies the sequential object recognition problem faced by a mobile robot searching for specific objects within a cluttered environment. In contrast to current state-of-the-art object recognition solutions which are evaluated on databases of static images, the system described in this paper employs an active strategy based on identifying potential objects using an attention mechanism and planning to obtain images of these objects from numerous viewpoints. We demonstrate the use of a bag-of-features technique for ranking potential objects, and show that this measure outperforms geometric matching for invariance across viewpoints. Our system implements informed visual search by prioritising map locations and re-examining promising locations first. Experimental results demonstrate that our system is a highly competent object recognition system that is capable of locating numerous challenging objects amongst distractors.

This paper introduces a novel colour-based affine co-variant region detector. Our algorithm is an extension of the maximally stable extremal region (MSER) to colour. The extension to colour is done by looking at successive time-steps of an agglomerative clustering of image pixels. The selection of time-steps is stabilised against intensity scalings and image blur by modelling the distribution of edge magnitudes. The algorithm contains a novel edge significance measure based on a Poisson image noise model, which we show performs better than the commonly used Euclidean distance. We compare our algorithm to the original MSER detector and a competing colour-based blob feature detector, and show through a repeatability test that our detector performs better. We also extend the state of the art in feature repeatability tests, by using scenes consisting of two planes where one is piecewise transparent. This new test is able to evaluate how stable a feature is against changing backgrounds.

This paper introduces an affine invariant shape descriptor for maximally stable extremal regions (MSER). Affine invariant feature descriptors are normally computed by sampling the original grey-scale image in an invariant frame defined from each detected feature, but we instead use only the shape of the detected MSER itself. This has the advantage that features can be reliably matched regardless of the appearance of the surroundings of the actual region. The descriptor is computed using the scale invariant feature transform (SIFT), with the resampled MSER binary mask as input. We also show that the original MSER detector can be modified to achieve better scale invariance by detecting MSERs in a scale pyramid. We make extensive comparisons of the proposed feature against a SIFT descriptor computed on grey-scale patches, and also explore the possibility of grouping the shape descriptors into pairs to incorporate more context. While the descriptor does not perform as well on planar scenes, we demonstrate various categories of full 3D scenes where it outperforms the SIFT descriptor computed on grey-scale patches. The shape descriptor is also shown to be more robust to changes in illumination. We show that a system can achieve the best performance under a range of imaging conditions by matching both the texture and shape descriptors.

A runtime system for implementation of image processing operations is presented. It is designed for working in a flexible and distributed environment related to the software architecture of a newly developed UAV system. The software architecture can be characterized at a coarse scale as a layered system, with a deliberative layer at the top, a reactive layer in the middle, and a processing layer at the bottom. At a finer scale each of the three levels is decomposed into sets of modules which communicate using CORBA, allowing system development and deployment on the UAV to be made in a highly flexible way. Image processing takes place in a dedicated module located in the process layer, and is the main focus of the paper. This module has been designed as a runtime system for data flow graphs, allowing various processing operations to be created online and on demand by the higher levels of the system. The runtime system is implemented in Java, which allows development and deployment to be made on a wide range of hardware/software configurations. Optimizations for particular hardware platforms have been made using Java's native interface.

In this paper, we present a new and efficient method to implement robust smoothing of low-level signal features: B-spline channel smoothing. This method consists of three steps: encoding of the signal features into channels, averaging of the channels, and decoding of the channels. We show that linear smoothing of channels is equivalent to robust smoothing of the signal features if we make use of quadratic B-splines to generate the channels. The linear decoding from B-spline channels allows the derivation of a robust error norm, which is very similar to Tukey's biweight error norm. We compare channel smoothing with three other robust smoothing techniques: nonlinear diffusion, bilateral filtering, and mean-shift filtering, both theoretically and on a 2D orientation-data smoothing task. Channel smoothing is found to be superior in four respects: It has a lower computational complexity, it is easy to implement, it chooses the global minimum error instead of the nearest local minimum, and it can also be used on nonlinear spaces, such as orientation space.