WO2021138991A1 - 一种弱监督可信代价传播的视差估计方法 - Google Patents

一种弱监督可信代价传播的视差估计方法 Download PDF

Info

Publication number
WO2021138991A1
WO2021138991A1 PCT/CN2020/077960 CN2020077960W WO2021138991A1 WO 2021138991 A1 WO2021138991 A1 WO 2021138991A1 CN 2020077960 W CN2020077960 W CN 2020077960W WO 2021138991 A1 WO2021138991 A1 WO 2021138991A1
Authority
WO
WIPO (PCT)
Prior art keywords
disparity
cost
map
credible
initial cost
Prior art date
Application number
PCT/CN2020/077960
Other languages
English (en)
French (fr)
Inventor
仲维
张宏
李豪杰
王智慧
刘日升
樊鑫
罗钟铉
李胜全
Original Assignee
大连理工大学
鹏城实验室
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 大连理工大学, 鹏城实验室 filed Critical 大连理工大学
Priority to US17/604,239 priority Critical patent/US11315273B2/en
Publication of WO2021138991A1 publication Critical patent/WO2021138991A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G06T7/55—Depth or shape recovery from multiple images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G06T7/55—Depth or shape recovery from multiple images
    • G06T7/593—Depth or shape recovery from multiple images from stereo images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761—Proximity, similarity or dissimilarity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/64—Three-dimensional objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20228—Disparity calculation for image-based rendering

Definitions

  • the invention belongs to the field of image processing and computer vision, relates to binocular disparity estimation, and uses a deep learning method to optimize the initial cost obtained by the traditional method. Specifically, it relates to a disparity estimation method for weakly supervised credible cost propagation.
  • Depth is the distance between the target and the observer as an important spatial information, which can be used to extract some important attributes (such as speed, three-dimensional position) and relationships (such as occlusion, out-of-bounds) of the target, which is of great significance for target tracking and behavior perception.
  • Binocular depth perception uses stereo matching to find the disparity between the corresponding points of the two images. According to the principle of triangulation, the depth value of the corresponding point in the scene is obtained with the internal and external parameters of the binocular camera.
  • the first category is traditional binocular disparity estimation methods, which are mainly based on geometric constraints, and are usually divided into matching cost calculation, cost aggregation, disparity estimation, and disparity optimization processes.
  • SGM Semi-Global Matching
  • the second category is disparity estimation methods based on deep learning, such as searching and matching using the features obtained by the network, using left and right consistency constraints to achieve disparity optimization and learning supervision.
  • researchers have proposed a method of using 3D convolution to realize the energy propagation process, which has further improved the interpretability of the network.
  • researchers have proposed an unsupervised method that uses left and right reconstruction consistency, and a transfer learning method based on the idea of domain adaptation.
  • the above data-driven learning method can obtain a feature model with stronger expressive ability. On the one hand, it can fully consider the semantic information, and on the other hand, it can learn the richer relationship characteristics between pixels. Therefore, the final disparity map is more than the result of the traditional method.
  • the third category is estimation methods that combine deep learning and traditional methods.
  • SGM algorithm using the network to automatically assign penalty coefficients to different scenes and different pixels can significantly improve the effect of SGM dense disparity estimation.
  • the literature uses the network to estimate the confidence of each pixel point, and carries out the energy propagation process according to the confidence.
  • these deep learning methods combined with traditional methods are more interpretable, they still do not fully utilize the respective advantages of the two types of methods. Therefore, compared with end-to-end learning methods, they fail to show advantages in accuracy, and are compared with sparseness.
  • the matching method does not show its advantages in terms of generalization ability and data dependence.
  • the present invention proposes a binocular disparity estimation method that combines deep learning with traditional methods, makes full use of the advantages of traditional methods and deep learning methods, and uses weakly supervised deep learning to optimize the rough initial cost obtained by traditional methods. , Obtain an accurate cost map, and solve a series of problems in the process of obtaining dense disparity maps, such as the difficulty of obtaining real disparity data labels, poor generalization ability across data sets, and mismatching of no texture and repeated texture regions.
  • the present invention aims to overcome the shortcomings of the prior art and provides a weakly supervised credible cost propagation disparity estimation method, that is, combining deep learning with traditional methods, using weakly supervised deep learning methods to optimize the initial cost map obtained by traditional methods, and effectively using traditional methods And the advantages of deep learning methods, to get a more accurate disparity map.
  • the specific plan includes the following steps:
  • the first step is to use the traditional feature matching method, that is, the non-deep learning method, to obtain a sparse and accurate initial cost map;
  • the second step is to transfer energy; use a three-dimensional convolutional network to optimize the initial cost map;
  • the third step is to perform disparity regression; convert the optimized initial cost map into a probability map, which is the probability that each pixel belongs to each disparity, and then obtain the sub-pixel disparity through Soft Argmax, and finally obtain a dense disparity map.
  • the present invention provides a binocular disparity estimation method combining traditional disparity estimation and deep learning. By combining and making full use of their respective advantages, it solves the problems of mismatching and difficulty in matching untextured regions in traditional methods. At the same time, the weakly-supervised credible cost propagation method avoids the problem of data label dependence in deep learning methods.
  • Figure 1 is the overall flow chart of the program
  • Figure 2 is the effect diagram of disparity estimation; (a) left image, (b) right image, (c) disparity image.
  • the present invention proposes a binocular disparity estimation scheme based on weakly supervised credible cost propagation ( Figure 1), the main steps are:
  • the first step is to use the traditional feature matching method to obtain a sparse and accurate cost map.
  • the present invention takes the matching method based on the Census feature as an example, and the specific steps are as follows:
  • HD( ⁇ ) represents the Hamming distance
  • CensusL and CensusR are the Census feature descriptors of the pixels on the left and right respectively.
  • the output of the cost calculation is a tensor of height (H) ⁇ width (W) ⁇ maximum disparity (D), that is, a cost map.
  • the second step is to spread energy.
  • the traditional energy propagation model is based on manual prior features and has limited description ability. It cannot effectively spread the energy in the same plane with too rich texture or the scene contains large areas of no texture and weak texture, and it is very easy to mismatch.
  • the present invention utilizes the powerful feature representation and context learning advantages of deep learning, and uses a three-dimensional convolutional network to optimize the cost map.
  • the input is an intermediate output of 1.2, which is a four-dimensional tensor composed of the similarity feature vector of each pixel with respect to each possible matching point.
  • the loss function corresponding to the energy propagation network is:
  • ⁇ is a point other than the sparse point set
  • the third step is to perform parallax regression.
  • the network is used to convert the similarity tensor into a probability tensor (that is, the probability that each pixel belongs to each disparity), and the sub-pixel disparity is obtained through Soft Argmax.
  • this method only needs to input the left image and the right image, and can output a sparse disparity map or a dense disparity map as needed.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Biology (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

本发明公开了一种弱监督可信代价传播的视差估计方法，利用深度学习方法对传统方法获得的初始代价进行优化。通过结合，充分利用各自优势，解决了传统方法中误匹配、无纹理区域匹配难等问题，同时弱监督的可信代价传播方法避免了深度学习方法数据标签依赖的问题。

Description

一种弱监督可信代价传播的视差估计方法 技术领域
本发明属于图像处理和计算机视觉领域，涉及双目视差估计，并利用深度学习方法对传统方法获得的初始代价进行优化。具体涉及一种弱监督可信代价传播的视差估计方法。
背景技术
深度即目标到观察者的距离作为一种重要的空间信息，可用来提取目标的一些重要属性(如速度、三维位置)和关系(如遮挡、越界)，对于目标跟踪及行为感知具有重要意义。双目深度感知利用立体匹配求出两幅图像对应点间的视差，根据三角测量原理，用双目相机内外参求得场景内相应点的深度值。现有的双目视差估计方法主要有三大类：第一类是传统双目视差估计方法，主要基于几何约束，通常分为匹配代价计算、代价聚合、视差估计、视差优化等过程。其中最经典的是SGM(半全局匹配)算法，其特点是每个像素点的代价用其对应位置八个方向路径上的像素代价做优化，在一定程度上解决了无纹理区域的匹配问题。传统方法对稀疏视差估计较准确，但估计稠密视差时局限性很大，如参数多，调参适配不同场景耗时耗力，特征设计难，能量传播不充分，特别是在镜面反射、低光、透明、无纹理区域仍存在较大缺陷。第二类是基于深度学习的视差估计方法，如用网络得到的特征进行搜索匹配，用左右一致性约束实现视差优化、学习的监督等。近两年研究者提出用3D卷积实现能量传播过程的方法，进一步提高了网络的可解释性。为解决数据依赖问题，研究者提出了利用左右重建一致性的无监督方法、基于域适应思想的迁移学习方法等。上述基于数据驱动的学习的方法，可得到表达能力更强的特征模型，一方面能充分考虑语义信息，另一方面能学到像素间的更丰富的关系特征，因此最终视差图较传统方法结果更准确、平滑，但存在着数据依赖强、场景泛化能力差的问题。第三类为深度学习与传统方法相结合的估计方法。如针对SGM算法，利用网络对不同场景、不同像素自动分配 惩罚系数，可显著提升SGM稠密视差估计效果。针对能量传播，文献利用网络估计出每个像素点的置信度，根据置信度进行能量传播过程。这些深度学习与传统方法结合的方法虽然可解释性更强，但仍未充分利用两类方法各自的优势，因此相对于端到端的学习方法未能在精度上体现出优势，而相较于稀疏匹配方法未在泛化能力及数据依赖程度上体现出优势。
本发明基于以上问题，提出一种将深度学习与传统方法相结合的双目视差估计法，充分利用传统方法与深度学习方法各自优势，利用弱监督深度学习对传统方法获取的粗糙初始代价进行优化，获得精确代价图，解决真实视差数据标签难获取、跨数据集泛化能力差、无纹理及重复纹理区域误匹配等稠密视差图获取过程中的一系列难题。
发明内容
本发明旨在克服现有技术的不足，提供了弱监督可信代价传播视差估计方法，即将深度学习与传统方法结合，利用弱监督深度学习方法优化传统方法获取的初始代价图，有效利用传统方法和深度学习方法各自的优势，得到更精确的视差图。
具体方案包括下列步骤：
一种基于弱监督可信代价传播的双目视差估计方法
第一步，采用传统特征匹配方法即非深度学习方法，得到稀疏精确的初始代价图；
第二步，进行能量传；采用三维卷积网络对初始代价图进行优化；
第三步，进行视差回归；利用优化后的初始代价图转化为概率图，所述概率图为每个像素属于每个视差的概率，再通过Soft Argmax得到亚像素视差，最终获得稠密视差图。
本发明的有益效果是：
本发明提出一种传统视差估计与深度学习结合的双目视差估计方法。通过结合，充分利用各自优势，解决了传统方法中误匹配、无纹理区域匹配难等问题，同时弱监督的可信代价传播方法避免了深度学习方法数据标签依赖的问题。
附图说明
图1为方案的整体流程图；
图2为视差估计的效果图；(a)左图，(b)右图，(c)视差图。
具体实施方式
为克服视差估计中，真实视差数据标签难获取、跨数据集泛化能力差、无纹理及重复纹理区域易误匹配等难题，本发明提出基于弱监督可信代价传播的双目视差估计方案(图1)，其主要步骤为：
第一步，采用传统特征匹配方法得到稀疏精确的代价图，本发明以基于Census特征的匹配方法为例，具体步骤如下：
1.1采用高斯滤波算法对输入图像进行降噪滤波处理。高斯滤波窗口权值由高斯函数(式1)决定。
Figure PCTCN2020077960-appb-000001
其中(x,y)是点坐标，σ是标准差。通过对高斯函数离散化，得到权值矩阵，即为高斯滤波器。
经过高斯滤波处理，能够有效的抑制噪声，平滑图像。防止噪声造成后续的匹配误差。
1.2对输入图像进行匹配，获得视差稀疏准确的初始代价图。。
进行初始代价计算。这里以基于Census特征的滑动窗口匹配为例描述匹配代价计算流程。
获取每个像素的Census特征描述子。利用滑动窗口，在扫描线上进行搜索，计算每个像素可能视差对应的代价(式2)：
Figure PCTCN2020077960-appb-000002
式中HD(·)表示汉明距离，CensusL、CensusR为分别为左图、右图像素的Census特征描述子。代价计算的输出为高(H)×宽(W)×最大视差(D)大小的张量， 即代价图。
第二步，进行能量传播。传统能量传播模型基于手工先验特征，描述能力有限，无法对同一平面内纹理过于丰富或场景包含大片无纹理、弱纹理区域进行有效能量传播，极易出现误匹配。本发明将利用深度学习强大的特征表征和上下文学习优势，采用三维卷积网络对代价图进行优化。输入为1.2的中间输出，即每个像素相对于每个可能匹配点的相似性特征向量组成的四维张量。能量传播网络对应的损失函数为：
Figure PCTCN2020077960-appb-000003
其中Ω为除稀疏点集之外的点，
Figure PCTCN2020077960-appb-000004
为左图上一点x，
Figure PCTCN2020077960-appb-000005
为根据右图与视差重构的左图上一点x。
第三步，进行视差回归。利用网络将相似性张量转换为概率张量(即每个像素属于每个视差的概率)，并通过Soft Argmax得到亚像素视差。实际应用中该方法仅需输入左图和右图，根据需要可输出稀疏视差图或稠密视差图。

Claims (2)

  1. 一种基于弱监督可信代价传播的双目视差估计方法，其特征在于，包括步骤如下：
    第一步，采用传统特征匹配方法即非深度学习方法，得到稀疏精确的初始代价图；
    第二步，进行能量传；采用三维卷积网络对初始代价图进行优化；
    第三步，进行视差回归；利用优化后的初始代价图转化为概率图，所述概率图为每个像素属于每个视差的概率，再通过Soft Argmax得到亚像素视差，最终获得稠密视差图。
  2. 根据权利要求1所述的基于弱监督可信代价传播的双目视差估计方法，其特征在于，第一步，采用传统特征匹配方法为基于Census特征的匹配方法，具体步骤如下：
    1)采用高斯滤波算法对输入图像进行降噪滤波处理，高斯滤波窗口权值由高斯函数，式(1)决定；
    Figure PCTCN2020077960-appb-100001
    其中(x,y)是点坐标，σ是标准差；通过对高斯函数离散化，得到权值矩阵，即为高斯滤波器；
    2)对输入图像进行匹配，获得视差稀疏准确的初始代价图；
    进行初始代价计算；获取每个像素的Census特征描述子，利用滑动窗口，在扫描线上进行搜索，计算每个像素可能视差对应的代价
    Figure PCTCN2020077960-appb-100002
    Figure PCTCN2020077960-appb-100003
    式中HD(·)表示汉明距离，CensusL、CensusR为分别为左图、右图像素的Census特征描述子；代价计算的输出为高(H)×宽(W)×最大视差(D)大小的张量，即初始代价图。
PCT/CN2020/077960 2020-01-10 2020-03-05 一种弱监督可信代价传播的视差估计方法 WO2021138991A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/604,239 US11315273B2 (en) 2020-01-10 2020-03-05 Disparity estimation method for weakly supervised trusted cost propagation

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010028299.7 2020-01-10
CN202010028299.7A CN111260711B (zh) 2020-01-10 2020-01-10 一种弱监督可信代价传播的视差估计方法

Publications (1)

Publication Number Publication Date
WO2021138991A1 true WO2021138991A1 (zh) 2021-07-15

Family

ID=70924122

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077960 WO2021138991A1 (zh) 2020-01-10 2020-03-05 一种弱监督可信代价传播的视差估计方法

Country Status (3)

Country Link
US (1) US11315273B2 (zh)
CN (1) CN111260711B (zh)
WO (1) WO2021138991A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103440653A (zh) * 2013-08-27 2013-12-11 北京航空航天大学 双目视觉立体匹配方法
WO2018211127A1 (en) * 2017-05-19 2018-11-22 Movidius Ltd. Methods, systems and apparatus to optimize pipeline execution
CN109241966A (zh) * 2018-08-22 2019-01-18 东北农业大学 一种植物叶片无损采集方法
CN109816710A (zh) * 2018-12-13 2019-05-28 中山大学 一种双目视觉系统高精度且无拖影的视差计算方法
CN110533712A (zh) * 2019-08-26 2019-12-03 北京工业大学 一种基于卷积神经网络的双目立体匹配方法

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101345303B1 (ko) * 2007-03-29 2013-12-27 삼성전자주식회사 스테레오 또는 다시점 영상의 입체감 조정 방법 및 장치
CN105956597A (zh) * 2016-05-04 2016-09-21 浙江大学 一种基于卷积神经网络的双目立体匹配方法
CN106355570B (zh) * 2016-10-21 2019-03-19 昆明理工大学 一种结合深度特征的双目立体视觉匹配方法
EP3514759A1 (en) * 2018-01-23 2019-07-24 Ricoh Company, Ltd. Image processing method, image processing apparatus, on-board device, moving body, and system
CN108416840B (zh) * 2018-03-14 2020-02-18 大连理工大学 一种基于单目相机的三维场景稠密重建方法
CN108416803B (zh) * 2018-03-14 2020-01-24 大连理工大学 一种基于深度神经网络的多信息融合的场景深度恢复方法
CN110310220A (zh) * 2018-03-27 2019-10-08 广东瑞图万方科技股份有限公司 一种半全局实时立体匹配方法
CN109146937A (zh) * 2018-08-22 2019-01-04 广东电网有限责任公司 一种基于深度学习的电力巡检影像密集匹配方法
CN111444744A (zh) * 2018-12-29 2020-07-24 北京市商汤科技开发有限公司 活体检测方法、装置以及存储介质
CN111383256B (zh) * 2018-12-29 2024-05-17 北京市商汤科技开发有限公司 图像处理方法、电子设备及计算机可读存储介质
CN109887019B (zh) * 2019-02-19 2022-05-24 北京市商汤科技开发有限公司 一种双目匹配方法及装置、设备和存储介质
CN110060290B (zh) * 2019-03-14 2021-06-04 中山大学 一种基于3d卷积神经网络的双目视差计算方法
CN110148181A (zh) * 2019-04-25 2019-08-20 青岛康特网络科技有限公司 一种通用双目立体匹配方法
CN110287964B (zh) * 2019-06-13 2021-08-03 浙江大华技术股份有限公司 一种立体匹配方法及装置
CN110400333B (zh) * 2019-07-26 2020-06-26 中国安全生产科学研究院 教练式双目立体视觉装置及高精度立体视觉图像获取方法
US11763433B2 (en) * 2019-11-14 2023-09-19 Samsung Electronics Co., Ltd. Depth image generation method and device

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103440653A (zh) * 2013-08-27 2013-12-11 北京航空航天大学 双目视觉立体匹配方法
WO2018211127A1 (en) * 2017-05-19 2018-11-22 Movidius Ltd. Methods, systems and apparatus to optimize pipeline execution
CN109241966A (zh) * 2018-08-22 2019-01-18 东北农业大学 一种植物叶片无损采集方法
CN109816710A (zh) * 2018-12-13 2019-05-28 中山大学 一种双目视觉系统高精度且无拖影的视差计算方法
CN110533712A (zh) * 2019-08-26 2019-12-03 北京工业大学 一种基于卷积神经网络的双目立体匹配方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JEONG SOMI; KIM SEUNGRYONG; HAM BUMSUB; SOHN KWANGHOON: "Convolutional cost aggregation for robust stereo matching", 2017 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP), IEEE, 17 September 2017 (2017-09-17), pages 2523 - 2527, XP033323030, DOI: 10.1109/ICIP.2017.8296737 *

Also Published As

Publication number Publication date
US20220092809A1 (en) 2022-03-24
US11315273B2 (en) 2022-04-26
CN111260711B (zh) 2021-08-10
CN111260711A (zh) 2020-06-09

Similar Documents

Publication Publication Date Title
CN109410321B (zh) 基于卷积神经网络的三维重建方法
Pham et al. Lcd: Learned cross-domain descriptors for 2d-3d matching
CN108665491B (zh) 一种基于局部参考点的快速点云配准方法
CN110827398B (zh) 基于深度神经网络的室内三维点云自动语义分割方法
Chen et al. Surface normals in the wild
CN109598754A (zh) 一种基于深度卷积网络的双目深度估计方法
CN109255833A (zh) 基于语义先验和渐进式优化宽基线致密三维场景重建方法
CN109766866B (zh) 一种基于三维重建的人脸特征点实时检测方法和检测系统
CN110517309A (zh) 一种基于卷积神经网络的单目深度信息获取方法
CN108305278A (zh) 一种orb-slam算法中图像匹配相关性改进方法
CN116310098A (zh) 一种基于注意力机制与可变卷积深度网络的多视图三维重建方法
Chen et al. Multi-dimensional cooperative network for stereo matching
Ubina et al. Intelligent underwater stereo camera design for fish metric estimation using reliable object matching
WO2022120988A1 (zh) 基于混合2d卷积和伪3d卷积的立体匹配方法
Wang et al. EFN6D: an efficient RGB-D fusion network for 6D pose estimation
CN114155406A (zh) 一种基于区域级特征融合的位姿估计方法
CN117351078A (zh) 基于形状先验的目标尺寸与6d姿态估计方法
WO2021138991A1 (zh) 一种弱监督可信代价传播的视差估计方法
CN116433904A (zh) 一种基于形状感知和像素卷积的跨模态rgb-d语义分割方法
Chen et al. Monocular image depth prediction without depth sensors: An unsupervised learning method
Li et al. Edge-aware neural implicit surface reconstruction
Gonzalez-Huitron et al. Jaccard distance as similarity measure for disparity map estimation
Chen et al. Marrying NeRF with Feature Matching for One-step Pose Estimation
Niu et al. Overview of image-based 3D reconstruction technology
Zhou et al. Geometry-Aware Network for Unsupervised Learning of Monocular Camera’s Ego-Motion

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20913046

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20913046

Country of ref document: EP

Kind code of ref document: A1