Using Synthetic Data to Estimate the True Error is Theoretically and Practically Doable

Generalizable and accurate stereo depth estimation is vital for 3D reconstruction, especially in surgery. Supervised learning methods obtain best performance however, limited ground truth data for surgical scenes limits generalizability. Self-supervised methods don't need ground truth, but suffer from scale ambiguity and incorrect disparity prediction due to inconsistency of photometric loss. This work proposes a two-phase training procedure that is generalizable and retains the high performance of supervised methods. It entails: (1) performing self-supervised representation learning of left and right views via masked image modelling (MIM) to learn generalizable semantic stereo features (2) utilizing the MIM pre-trained model to learn robust depth representation via supervised learning for disparity estimation on synthetic data only. To improve stereo representations learnt via MIM, perceptual loss terms are introduced, which improve the model's stereo representations learnt by explicitly encouraging the learning of higher scene-level features. Qualitative and quantitative performance evaluation on surgical and natural scenes shows that the approach achieves sub-millimetre accuracy and lowest errors respectively, setting a new state-of-the-art. Despite not training on surgical nor natural scene data for disparityestimation.

This paper is about an interesting application of the Frequency Modulated Continuous Wave (FMCW) radar system for the traffic surveillance.By introducing a novel monitoring mode, it is designed to provide synchronously the velocity measurements of all the vehicles within the radar footprint.To achieve a precise estimation of the real speeds of the moving vehicles, a Joint Maximum Contrast and Map Drift (JMCMD) algorithm using the autofocus technique is proposed in this paper.Those autofocus techniques originated from synthetic aperture radar data processing are analyzed and coupled here to form a two-step algorithm.Firstly, the maximum contrast method is used for a coarse estimation to get the rough velocity of the target.Then, the more accurate velocity of the target is obtained by using the map drift algorithm.The simulation results illustrate the errors of the measurement accuracy are less than 1km/h, which satisfy the practical requirements.

We believe that a wide range of physical processes conspire to shape the observed galaxy population, but we remain unsure of their detailed interactions. The semi-analytic model (SAM) of galaxy formation uses multidimensional parametrizations of the physical processes of galaxy formation and provides a tool to constrain these underlying physical interactions. Because of the high dimensionality, the parametric problem of galaxy formation may be profitably tackled with a Bayesian-inference-based approach, which allows one to constrain theory with data in a statistically rigorous way. In this paper, we develop a SAM in the framework of Bayesian inference. We show that, with a parallel implementation of an advanced Markov chain Monte Carlo algorithm, it is now possible to rigorously sample the posterior distribution of the high-dimensional parameter space of typical SAMs. As an example, we characterize galaxy formation in the current Λ cold dark matter cosmology using the stellar mass function of galaxies as an observational constraint. We find that the posterior probability distribution is both topologically complex and degenerate in some important model parameters, suggesting that thorough explorations of the parameter space are needed to understand the models. We also demonstrate that because of the model degeneracy, adopting a narrow prior strongly restricts the model. Therefore, the inferences based on SAMs are conditional to the model adopted. Using synthetic data to mimic systematic errors in the stellar mass function, we demonstrate that an accurate observational error model is essential to meaningful inference.

SUMMARY Gravity field solutions produced using data from the Gravity Recovery and Climate Experiment (GRACE) satellite mission are subject to errors that increase as a function of increasing spatial resolution. Two commonly used techniques to improve the signal-to-noise ratio in the gravity field solutions are post-processing, via spectral filters, and regularization, which occurs within the least-squares inversion process used to create the solutions. One advantage of post-processing methods is the ability to easily estimate the signal loss resulting from the application of the spectral filter by applying the filter to synthetic gravity field coefficients derived from models of mass variation. This is a critical step in the construction of an accurate error budget. Estimating the amount of signal loss due to regularization, however, requires the execution of the full gravity field determination process to create synthetic instrument data; this leads to a significant cost in computation and expertise relative to post-processing techniques, and inhibits the rapid development of optimal regularization weighting schemes. Thus, while a number of studies have quantified the effects of spectral filtering, signal modification in regularized GRACE gravity field solutions has not yet been estimated. In this study, we examine the effect of one regularization method. First, we demonstrate that regularization can in fact be performed as a post-processing step if the solution covariance matrix is available. Regularization then is applied as a post-processing step to unconstrained solutions from the Center for Space Research (CSR), using weights reported by the Centre National d'Etudes Spatiales/Groupe de Recherches de geodesie spatiale (CNES/GRGS). After regularization, the power spectra of the CSR solutions agree well with those of the CNES/GRGS solutions. Finally, regularization is performed on synthetic gravity field solutions derived from a land surface model, revealing that in some locations significant signal loss can result from regularization. This signal loss is similar in magnitude to estimated signal loss in post-filtered solutions. End-users of GRACE data can use this method to improve the error budgets of GRACE time-series, or to restore the power lost through regularization using a scaling technique.

The uncertainty of a PIV displacement field is estimated using a generic post-processing method based on statistical analysis of the correlation process using differences in the intensity pattern in the two images. First the second image is dewarped back onto the first one using the computed displacement field which provides two almost perfectly matching images. Differences are analyzed regarding the effect of shifting the peak of the correlation function. A relationship is derived between the standard deviation of intensity differences in each interrogation window and the expected asymmetry of the correlation peak, which is then converted to the uncertainty of a displacement vector. This procedure is tested with synthetic data for various types of noise and experimental conditions (pixel noise, out-of-plane motion, seeding density, particle image size, etc) and is shown to provide an accurate estimate of the true error.