Self-supervised monocular depth estimation (MDE) models universally suffer from the notorious edge-fattening issue. Triplet loss, as a widespread metric learning strategy, has largely succeeded in many computer vision applications. In this paper, we redesign patch-based triplet loss MDE to alleviate ubiquitous We show two drawbacks of raw and demonstrate our problem-driven redesigns. First, present min. operator based strategy applied all negative samples, prevent well-performing negatives...

In recent years, Deep Learning based method for 3-Dimension (3D) geometry perception tasks, like dense depth recovery, optical flow estimation and ego-motion estimation, is attracting significant attention. Inspired by advances in unsupervised strategies to learning from video datasets, we present a reasonable combination of constrains finer architecture, used estimation. Specifically, introduce our effective neural networks Depth-Net (for monocular estimation) Pose-Net estimation), which...

Unsupervised learning-based scene perception has recently become an important research direction. Most unsupervised methods for tasks (e.g., dense depth recovery and ego-motion estimation) train the convolutional network via minimizing photometric error of images, achieving very impressive results. Since supervision signal is weaker than ground truth data, existing generally do poorly inaccurate pose estimation high-resolution map generation. To this end, we present architecture based on...

Bioinspired polarized skylight navigation, which can be used in unfamiliar territories, is an important alternative autonomous navigation technique the absence of Global Navigation Satellite System (GNSS). However, polarization pattern night environment with noise effects and model uncertainties a less explored area. Although several decades have passed since first publication about moonlit sky, usefulness nocturnal only sporadic previous researches. This study demonstrates that light...

Although the image object detection technology developed quickly, in videos has special conditions such as motion blur, video defocus, part occlusion, rare posture, and needs to keep target result consistent time. It cannot apply models directly on videos. In this paper, we proposed method based YOLO-v3 together with FlowNet 2.0 optical flow extraction network make full use of information frames. The Flow-guided Partial Warp operates at feature map level more accurate. order solve blur...

The multilayer bucketing screener has proved to be effective tune the sparse Speeded-Up Robust Features (SURF) flow. It not only reduces influence of illumination changes but also makes distribution optical flow more uniform. Based on SURF with screener, a complete scheme for ego-motion estimation in monocular vision systems is proposed. Taking advantage two-view obtain relative scale, we use ground plane calculate absolute scale. Random sample consensus-based outlier rejection schemes are...

Different from single image dehazing, polarized images can record richer scene information, and thus images-based dehazing has attracted tremendous attention recently. Existing methods usually propose some priors or assumptions to calculate the degree of polarization angle for dehazing. Although these have made significant progress, are not always reliable in real scene, which limits their performance. In addition, most existing only consider pixel-level difference information ignore...

This paper presents an approach capable of recovering the trajectory from image sequence a single camera based on sparse SURF flow with multi-layer bucketing screener. The main novelty proposed is that we designed screener to screen good features detected by algorithm. scheme firstly divides into several buckets. For each bucket, both mismatched and redundant feature pairs will be rejected so accuracy improved, also maximal number selected kept in order reduce computational complexity....

A variety of techniques are currently presented for querying and mining time series data based on different kinds representations similarity measures. These mainly focus the numerical characteristics sensitive to changes series. However, we find that generally contain curves sharing some set visual features. offer a deeper understanding data, open up potential new technique analysis. Particularly beneficial from recent advances in deep neural networks (DNNs), features can be automatically...

The outliers caused by noise and mismatching severely restrict the precision of visual odometry. Moreover, dynamic environment is also a crucial element that decreases robustness systems. This paper presents robust stereo odometry decoupled ego-motion estimation based on probabilistic matches rejecting objects through motion segmentation. Fast ZNCC method, local sum table partition upper bound schemes, presented for selecting while keeping run-time efficiency. selection multi-correspondences...

The task of generating comments for images has been studied by many scholars. Existing methods, such as sentence generation based on recurrent neural network (RNN), have poor variability due to their structural limitations. In addition, the generated comment is so stiff that lacks emotional color and value orientation human language. order make image more open meet needs emotion control, we designed a style transfer module pictures. This paper mainly completes following work: (1) Imitate...

Ionosphere occultation inversion technique is an important means of the ionosphere exploration. Researches on its accuracy and anti-disturbance performance contribute to improve retrieval results, which also useful assess applied conditions different methods. In this paper, we firstly analyses three kinds methods for electron density in detail, including Abel integral method, "Onion-layered" method least squares method. Then, put forward a smoothing way by averaging fitting functions with...

The popular mathematical and simulation methods both have some shortage in evaluating the stationary operational availability. former assumes that item demand is independent of operating systems, which can introduce a serious underestimation latter order to get acceptable confidence level requires large trials under pre-assigned scenario. This paper addresses issue determining availability basing on models. Not only system passivation operation mission, but also other influencing factors,...

This paper presents a novel approach to compute sparse optical flow field, which is different from the traditional feature matching methods such as SURF. The consists mainly of three parts. First, improved PatchMatch that tailored computation used generate NNF quickly. And smoothed using threshold filtering rather than global optimization for low complexity. Next, we use superpixel method segment and choose representative each statistical filtering. Finally, outliers are mistaken inliers in...

Due to the lack of ground-truth, nearly all self-supervised depth estimation methods use reprojection errors as main supervised signal. We find that it fails train network with pixels in some special regions, such textureless, boundary, occlusion and motion regions. The reason is error-based loss produces an unexpected deviation because irregular illumination change wrong matching pixels. As a result, we propose novel geometric-inspired calibration mechanism (GCM) contains two parts rectify...