CN108447082A - A kind of objective matching process based on combination learning Keypoint detector - Google Patents
A kind of objective matching process based on combination learning Keypoint detector Download PDFInfo
- Publication number
- CN108447082A CN108447082A CN201810215020.9A CN201810215020A CN108447082A CN 108447082 A CN108447082 A CN 108447082A CN 201810215020 A CN201810215020 A CN 201810215020A CN 108447082 A CN108447082 A CN 108447082A
- Authority
- CN
- China
- Prior art keywords
- key point
- loss
- image
- score
- correspondence
- Prior art date
- Legal status (The legal status 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 status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
- G06T7/344—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods involving models
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
A kind of objective matching process based on combination learning Keypoint detector proposed in the present invention, main contents include:Faster region convolutional neural networks, training, combined optimization, its process is, given two depth images with some attitude disturbances, two groups of suggestions are generated to each image, then carry out projection three-dimensional suggestion using known images of gestures, to establish face and negative pair, by these to passing to comparison loss, attempt to make the characteristic distance between face to minimize and make it is negative to the distance between maximize, introduce new fractional loss, adjustment region suggests the parameter of network, in the suggestion of the Area generation balloon score of depth map, find correspondence, to complete to match.The present invention proposes a kind of new sample level, can be with the corresponding label of in-time generatin part patch, while improving the accuracy of target correspondence, and qualitative and quantitative improvement has been carried out to the prior art.
Description
Technical field
The present invention relates to object matching fields, more particularly, to a kind of three-dimensional based on combination learning Keypoint detector
Target matching method.
Background technology
With the development of human society, object matching technology has become modern society's field of information processing, especially
An especially important technology in image processing field.In image registration, camera calibration, target identification and image retrieval etc.
In computer vision application, the key point that same object or scene are correctly matched from two images is a basis and important
Work.Object matching technology be widely used in daily life, production, military activity, aerospace detection, target identification with
In a series of field such as tracking, robot vision.Such as in field of image search, by user's input picture and database
Image carries out matching search, and user can be helped to search the relevant information of target image content.In robot vision field, machine
People is compared with stored database by identifying target object, can obtain object information, to preferably complete to refer to
Determine task.The overall situation indicates it is common method the problems such as solving matching, retrieval, Attitude estimation and registration, can be with directly side
Formula is trained in a manner of end to end, and disadvantage is susceptible to that the blocking of a large amount of confusions, partial view or scene block.
The present invention proposes a kind of objective matching process based on combination learning Keypoint detector, gives two tools
The depth image for having some attitude disturbances is generated two groups of suggestions to each image, is then projected using known images of gestures
Three-dimensional is suggested, to establish face and negative pair, by these to passing to comparison loss, it is intended to make characteristic distance between face most
Smallization and make it is negative to the distance between maximize, introduce new fractional loss, adjustment region suggests the parameter of network, in depth
The suggestion of the Area generation balloon score of figure, finds correspondence, to complete to match.The present invention proposes a kind of new sampling
Layer, can be with the corresponding label of in-time generatin part patch, while improving the accuracy of target correspondence, is carried out to the prior art
Qualitative and quantitative improvement.
Invention content
For be easy to be blocked by a large amount of confusions, partial view or scene are blocked the problem of, it is an object of the invention to
A kind of objective matching process based on combination learning Keypoint detector is provided, giving two has some attitude disturbances
Depth image generates two groups of suggestions to each image, then carrys out projection three-dimensional suggestion using known images of gestures, to establish
Face and negative pair, by these to passing to comparison loss, it is intended to so that the characteristic distance between face is minimized and make to bear between
Distance maximize, introduce new fractional loss, adjustment region suggests the parameter of network, in the Area generation high score of depth map
Several suggestions, finds correspondence, to complete to match.
To solve the above problems, the present invention provides a kind of objective match party based on combination learning Keypoint detector
Method, main contents include:
(1) faster region convolutional neural networks;
(2) training;
(3) combined optimization.
Wherein, the combination learning Keypoint detector uses modified faster region convolutional neural networks (R-
CNN) carry out study-leading process;Specifically, two depth images with some attitude disturbances are given, first to each image
Generate two groups of suggestions;Then, carry out projection three-dimensional suggestion using known images of gestures, to establish face and negative pair;In three dimensions
Suggestion with small distance is considered as correspondence, therefore is marked as just;Then by these to passing to comparison loss, with
Attempt to make the characteristic distance between face to minimize and make it is negative to the distance between maximize;In addition, introducing a kind of new point
Number loss, it has adjusted the parameter that network (RPN) is suggested in the faster regions R-CNN, in the Area generation balloon score of depth map
Suggestion, correspondence can be found always.
Wherein, the faster region convolutional neural networks, using faster R-CNN as the company with shared weight
A part for body Model;Liang Ge branches are all connected to a layer for being responsible for searching correspondence, referred to as sample level;Use comparison
Loss indicates to train, and there is the fractional loss in trained critical point detection stage in each branch.
Wherein, the training needs pairs of depth image { I for training pattern0,I1, each depth image tool
There is its camera posture information { g0,g1And intrinsic camera parameter C;These can be by from multiple viewpoint renders three-dimensional models or making
With RGB-D video sequences be registrated frame;In order to by depth image { I0,I1By network, their depth value is normalized first
For RGB ranges, and single channel is copied into triple channel image;Input g0,g1, the rest part and depth image of C, value with
Rice is unit D0,D1It is directly delivered to sample level.
Further, the depth image, for each depth image, region suggests that network (RPN) can generate one group
Score and area-of-interest (RoIs), using their barycenter as key point position.
Further, the area-of-interest, each RoI also determine the space of the feature calculation for current key point
Range, and after the layer of the ponds RoI, obtain the expression of each key point;Retain highest t key points according to their score,
And establish key pointCorresponding a pair of of the depth images of wherein m={ 0,1 }, It is the two-dimensional coordinate on the plane of delineation,It is the score for indicating key point significance degree,It is corresponding spy
Sign vector;Then sample level receives key point barycenter and its feature group { x from two images0,x1,f0,f1};Finally determine two
Correspondence between the key point of width image.
Further, the correspondence between the key point of the determination two images, barycenter project to three-dimensional first
In space;For each key pointIt is found according to Euclidean distance immediate in three dimensionsAnd form n-th pair
Feature If distance is less than a small threshold value, it is marked as faceOtherwise will
It is considered as negative pair
Wherein, the combined optimization, combination learning viewpoint invariant representation and Keypoint detector;It introduces more following
Business loss function:
Wherein, LcIt is the comparison loss slightly changed, to key point to carrying out operation and being optimized to expression,
It is to use key point score with the fractional loss component of optimized detector, l ' is set of the feature to F ', λcAnd λsIt is weight ginseng
Number.
Further, the feature pair is indicated since these features define one group to F ' using mark n
N-th of feature pairComparison loss is defined as:
Wherein, v indicates back gauge, NposAnd NnegIt is positive and negative couple of quantity (N=N respectivelypos+Nneg);Each classification is to loss
Contribution be all to be normalized according to its group, to explain face and the negative imbalance between;Fractional loss defines
For:
Wherein,It is to come from image ImI-th of key point label, value depending on key point belong to face or
Negative pair, and γ is regularization parameter;Due to these to be by from each image choose a key point formed, and
Each key point can only belong to one it is right, then | l0|=| l1|=N.
Further, the purpose of the fractional loss, fractional loss is the correspondence maximized between two views
Quantity;Optimization generates more positive key points as much as possible, and maximizes their score;If the score generated is very low,
Only consider face and they are punished;Loss is standardized by the quantity of positive number, however, it is possible to which γ is excellent to adjust
Change keypoint quantity and optimizes the tradeoff between score;Frame allows to adjust by adjusting three-dimensional distance threshold value in sample level
Tradeoff during training between matching times and positioning accuracy;During backpropagation, by being stored during forward direction is transmitted
Simultaneously propose to realize function backward in layer in region, the gradual change of each key point is transmitted to suitable in gradual change mapping in their position
Work as position;For fractional loss, gradient predicts the convolutional layer of score by being responsible for.
Description of the drawings
Fig. 1 is a kind of system flow of the objective matching process based on combination learning Keypoint detector of the present invention
Figure.
Fig. 2 is a kind of faster region of the objective matching process based on combination learning Keypoint detector of the present invention
Convolutional neural networks framework.
Fig. 3 is a kind of present invention network training phase of the objective matching process based on combination learning Keypoint detector
Between gradient reverse propagate.
Specific implementation mode
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
It mutually combines, invention is further described in detail in the following with reference to the drawings and specific embodiments.
Fig. 1 is a kind of system flow of the objective matching process based on combination learning Keypoint detector of the present invention
Figure.Include mainly faster region convolutional neural networks, training, combined optimization.
Combination learning Keypoint detector is guided using modified faster region convolutional neural networks (R-CNN)
Learning process;Specifically, two depth images with some attitude disturbances are given, generating two to each image first sets up
View;Then, carry out projection three-dimensional suggestion using known images of gestures, to establish face and negative pair;There is small distance in three dimensions
Suggestion be considered as correspondence, therefore be marked as just;Then by these to passing to comparison loss, to attempt to make face
Between characteristic distance minimize and make it is negative to the distance between maximize;In addition, introducing a kind of new fractional loss, it is adjusted
The parameter of network (RPN) is suggested in the whole faster regions R-CNN, can be with in the suggestion of the Area generation balloon score of depth map
Always correspondence is found.
Combined optimization, combination learning viewpoint invariant representation and Keypoint detector;Introduce following multitask loss function:
Wherein, LcIt is the comparison loss slightly changed, to key point to carrying out operation and being optimized to expression,
It is to use key point score with the fractional loss component of optimized detector, l ' is set of the feature to F ', λcAnd λsIt is weight ginseng
Number.
Since these features define one group to F ', n-th of feature pair is indicated using mark nComparison
Loss is defined as:
Wherein, v indicates back gauge, NposAnd NnegIt is positive and negative couple of quantity (N=N respectivelypos+Nneg);Each classification is to loss
Contribution be all to be normalized according to its group, to explain face and the negative imbalance between;Fractional loss defines
For:
Wherein,It is to come from image ImI-th of key point label, value depending on key point belong to face or
Negative pair, and γ is regularization parameter;Due to these to be by from each image choose a key point formed, and
Each key point can only belong to one it is right, then | l0|=| l1|=N.
Fig. 2 is a kind of faster region of the objective matching process based on combination learning Keypoint detector of the present invention
Convolutional neural networks framework.Using faster R-CNN as a part for the connected with auxilliary crankshuft model with shared weight;Liang Ge branches are all
It is connected to a layer for being responsible for searching correspondence, referred to as sample level;It is indicated to train using comparison loss, and each divided
Zhi Douyou trains the fractional loss in critical point detection stage.
For training pattern, pairs of depth image { I is needed0,I1, each depth image has its camera posture information
{g0,g1And intrinsic camera parameter C;These can by from multiple viewpoint renders three-dimensional models or using RGB-D video sequences with
It is registrated frame;In order to by depth image { I0,I1By network, their depth value is normalized to RGB ranges first, and will be single
A channel is copied into triple channel image;Input g0,g1, the rest part and depth image of C, value is using rice as unit D0,D1Directly
It is transmitted to sample level.
For each depth image, region suggests that network (RPN) can generate a number of components and area-of-interest (RoIs), makes
Use their barycenter as key point position.
Each RoI also determines the spatial dimension of the feature calculation for current key point, and after the layer of the ponds RoI,
Obtain the expression of each key point;Retain highest t key points according to their score, and establishes key point Corresponding a pair of of the depth images of wherein m={ 0,1 },It is that image is flat
Two-dimensional coordinate on face,It is the score for indicating key point significance degree,It is corresponding feature vector;Then sample level from
Key point barycenter and its feature group { x are received in two images0,x1,f0,f1};Between the key point for finally determining two images
Correspondence.
Barycenter projects in three dimensions first;For each key pointThree-dimensional space is found according to Euclidean distance
Between in it is immediateAnd form n-th pair of featureIf distance is less than a small threshold value, it is marked as
Face Otherwise it is regarded as negative pair
Fig. 3 is a kind of present invention network training phase of the objective matching process based on combination learning Keypoint detector
Between gradient reverse propagate.The purpose of fractional loss is to maximize the quantity of the correspondence between two views;As much as possible
Optimization generates more positive key points, and maximizes their score;If the score generated is very low, face is only considered and to it
Punished;Loss standardized by the quantity of positive number, however, it is possible to γ come adjust optimize keypoint quantity with it is excellent
Change the tradeoff between score;Frame allows by adjusting the three-dimensional distance threshold value in sample level come matching times during adjusting training
Tradeoff between positioning accuracy;During backpropagation, by storing their position during forward direction is transmitted and in area
Function backward is realized in domain proposal layer, and the gradual change of each key point is transmitted to the appropriate location in gradual change mapping;For score
Loss, gradient predict the convolutional layer of score by being responsible for.
For those skilled in the art, the present invention is not limited to the details of above-described embodiment, in the essence without departing substantially from the present invention
In the case of refreshing and range, the present invention can be realized in other specific forms.In addition, those skilled in the art can be to this hair
Bright to carry out various modification and variations without departing from the spirit and scope of the present invention, these improvements and modifications also should be regarded as the present invention's
Protection domain.Therefore, the following claims are intended to be interpreted as including preferred embodiment and falls into all changes of the scope of the invention
More and change.
Claims (10)
1. a kind of objective matching process based on combination learning Keypoint detector, which is characterized in that main includes faster
Region convolutional neural networks (one);Training (two);Combined optimization (three).
2. based on the combination learning Keypoint detector described in claims 1, which is characterized in that using modified faster
Region convolutional neural networks (R-CNN) carry out study-leading process;Specifically, two depth with some attitude disturbances are given
Image generates two groups of suggestions to each image first;Then, carry out projection three-dimensional suggestion using known images of gestures, to establish
Face and negative pair;The suggestion with small distance is considered as correspondence in three dimensions, therefore is marked as just;Then by these
To passing to comparison loss, with attempt to make the characteristic distance between face to minimize and make it is negative to the distance between maximization;This
Outside, a kind of new fractional loss is introduced, it has adjusted the parameter that network (RPN) is suggested in the faster regions R-CNN, in depth
The suggestion of the Area generation balloon score of figure, can find correspondence always.
3. based on the faster region convolutional neural networks (one) described in claims 1, which is characterized in that use faster R-
A parts of the CNN as the connected with auxilliary crankshuft model with shared weight;Liang Ge branches are all connected to one and are responsible for searching correspondence
Layer, referred to as sample level;It is indicated using comparison loss to train, and there be point in trained critical point detection stage in each branch
Number loss.
4. based on the training (two) described in claims 1, which is characterized in that for training pattern, need pairs of depth map
As { I0,I1, each depth image has its camera posture information { g0,g1And intrinsic camera parameter C;These can by from
Multiple viewpoint renders three-dimensional models use RGB-D video sequences and are registrated frame;In order to by depth image { I0,I1By network,
Their depth value is normalized to RGB ranges first, and single channel is copied into triple channel image;Input g0,g1, C its
Remaining part point and depth image, value is using rice as unit D0,D1It is directly delivered to sample level.
5. based on the depth image described in claims 4, which is characterized in that for each depth image, network is suggested in region
(RPN) number of components and area-of-interest (RoIs) can be generated, using their barycenter as key point position.
6. based on the area-of-interest described in claims 5, which is characterized in that each RoI, which is also determined, is used for current key point
Feature calculation spatial dimension, and after the layer of the ponds RoI, obtain the expression of each key point;According to their score
Retain highest t key points, and establishes key pointWherein m={ 0,1 } is right
A pair of of depth image is answered,It is the two-dimensional coordinate on the plane of delineation,It is point for indicating key point significance degree
Number,It is corresponding feature vector;Then sample level receives key point barycenter and its feature group { x from two images0,x1,
f0,f1};Finally determine the correspondence between the key point of two images.
7. based on the correspondence between the key point of the determination two images described in claims 6, which is characterized in that barycenter
It projects in three dimensions first;For each key pointIt is found according to Euclidean distance immediate in three dimensionsAnd form n-th couple of feature F 'n=(fi 0,fj 1);If distance is less than a small threshold value, it is marked as face Otherwise it is regarded as negative pair
8. based on the combined optimization (three) described in claims 1, which is characterized in that combination learning viewpoint invariant representation and pass
Key spot detector;Introduce following multitask loss function:
Wherein, LcIt is the comparison loss slightly changed, to key point to carrying out operation and being optimized to expression,It is to make
With key point score with the fractional loss component of optimized detector, l ' is set of the feature to F ', λcAnd λsIt is weight parameter.
9. based on the feature pair described in claims 8, which is characterized in that since these features define one group to F ',
N-th of feature pair is indicated using mark nComparison loss is defined as:
Wherein, v indicates back gauge, NposAnd NnegIt is positive and negative couple of quantity (N=N respectivelypos+Nneg);Each tribute of the classification to loss
Offering is normalized according to its group, to explain face and the negative imbalance between;Fractional loss is defined as:
Wherein,It is to come from image ImI-th of key point label, value belongs to face or negative depending on key point
It is right, and γ is regularization parameter;Due to these to be by from each image choose a key point is formed, and often
A key point can only belong to one it is right, then | l0|=| l1|=N.
10. based on the fractional loss described in claims 8, which is characterized in that the purpose of fractional loss is to maximize two to regard
The quantity of correspondence between figure;Optimization generates more positive key points as much as possible, and maximizes their score;If
The score of generation is very low, only considers face and punishes them;Loss is standardized by the quantity of positive number, but
It is that adjust optimization keypoint quantity and the tradeoff between score can be optimized with γ;Frame allows by adjusting in sample level
Three-dimensional distance threshold value is come the tradeoff during adjusting training between matching times and positioning accuracy;During backpropagation, pass through
Their position is stored during positive transmit and proposes to realize function backward in layer in region, and the gradual change of each key point is passed
It is delivered to the appropriate location in gradual change mapping;For fractional loss, gradient predicts the convolutional layer of score by being responsible for.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810215020.9A CN108447082A (en) | 2018-03-15 | 2018-03-15 | A kind of objective matching process based on combination learning Keypoint detector |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810215020.9A CN108447082A (en) | 2018-03-15 | 2018-03-15 | A kind of objective matching process based on combination learning Keypoint detector |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108447082A true CN108447082A (en) | 2018-08-24 |
Family
ID=63194604
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810215020.9A Withdrawn CN108447082A (en) | 2018-03-15 | 2018-03-15 | A kind of objective matching process based on combination learning Keypoint detector |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108447082A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110533639A (en) * | 2019-08-02 | 2019-12-03 | 杭州依图医疗技术有限公司 | A kind of key independent positioning method and device |
CN110743818A (en) * | 2019-11-29 | 2020-02-04 | 苏州嘉诺环境工程有限公司 | Garbage sorting system and garbage sorting method based on vision and deep learning |
WO2020063986A1 (en) * | 2018-09-30 | 2020-04-02 | 先临三维科技股份有限公司 | Method and apparatus for generating three-dimensional model, device, and storage medium |
CN112631120A (en) * | 2019-10-09 | 2021-04-09 | Oppo广东移动通信有限公司 | PID control method, device and video coding and decoding system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
CN107145852A (en) * | 2017-04-28 | 2017-09-08 | 深圳市唯特视科技有限公司 | A kind of character recognition method based on homologous cosine losses function |
-
2018
- 2018-03-15 CN CN201810215020.9A patent/CN108447082A/en not_active Withdrawn
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
CN107145852A (en) * | 2017-04-28 | 2017-09-08 | 深圳市唯特视科技有限公司 | A kind of character recognition method based on homologous cosine losses function |
Non-Patent Citations (1)
Title |
---|
GEORGIOS GEORGAKIS: "End-to-end learning of keypoint detector and descriptor for pose invariant 3D matching", 《ARXIV:1802.07869V1》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020063986A1 (en) * | 2018-09-30 | 2020-04-02 | 先临三维科技股份有限公司 | Method and apparatus for generating three-dimensional model, device, and storage medium |
US11978157B2 (en) | 2018-09-30 | 2024-05-07 | Shining 3D Tech Co., Ltd. | Method and apparatus for generating three-dimensional model, device, and storage medium |
CN110533639A (en) * | 2019-08-02 | 2019-12-03 | 杭州依图医疗技术有限公司 | A kind of key independent positioning method and device |
CN112631120A (en) * | 2019-10-09 | 2021-04-09 | Oppo广东移动通信有限公司 | PID control method, device and video coding and decoding system |
CN112631120B (en) * | 2019-10-09 | 2022-05-17 | Oppo广东移动通信有限公司 | PID control method, device and video coding and decoding system |
CN110743818A (en) * | 2019-11-29 | 2020-02-04 | 苏州嘉诺环境工程有限公司 | Garbage sorting system and garbage sorting method based on vision and deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Li et al. | Deepim: Deep iterative matching for 6d pose estimation | |
Zakharov et al. | Dpod: 6d pose object detector and refiner | |
CN110021051B (en) | Human image generation method based on generation of confrontation network through text guidance | |
Abdelrahman et al. | L2cs-net: Fine-grained gaze estimation in unconstrained environments | |
Gilani et al. | Deep, dense and accurate 3D face correspondence for generating population specific deformable models | |
CN108447082A (en) | A kind of objective matching process based on combination learning Keypoint detector | |
WO2021012484A1 (en) | Deep learning-based target tracking method and apparatus, and computer readable storage medium | |
Quattoni et al. | Hidden-state conditional random fields | |
Kuhnke et al. | Deep head pose estimation using synthetic images and partial adversarial domain adaption for continuous label spaces | |
Cheng et al. | Robust visual localization in dynamic environments based on sparse motion removal | |
CN113610126A (en) | Label-free knowledge distillation method based on multi-target detection model and storage medium | |
Guo et al. | A generalized and robust method towards practical gaze estimation on smart phone | |
CN108027878A (en) | Method for face alignment | |
Wang et al. | Graph-based multiprototype competitive learning and its applications | |
Zhang et al. | Transductive Learning Via Improved Geodesic Sampling. | |
Scheffler et al. | Joint adaptive colour modelling and skin, hair and clothing segmentation using coherent probabilistic index maps | |
CN110119768B (en) | Visual information fusion system and method for vehicle positioning | |
Ott et al. | Joint classification and trajectory regression of online handwriting using a multi-task learning approach | |
CN108564022A (en) | A kind of more personage's pose detection methods based on positioning classification Recurrent networks | |
Huang et al. | Centroid networks for few-shot clustering and unsupervised few-shot classification | |
Wang et al. | Joint head pose and facial landmark regression from depth images | |
Zhou et al. | MTCNet: Multi-task collaboration network for rotation-invariance face detection | |
Gattone et al. | A shape distance based on the Fisher–Rao metric and its application for shapes clustering | |
CN116030519A (en) | Learning attention detection and assessment method for live broadcast teaching platform | |
Chen et al. | Hierarchical posture representation for robust action recognition |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180824 |
|
WW01 | Invention patent application withdrawn after publication |