CN108470107A - A kind of apartment three-dimension object layout generation method based on probabilistic model - Google Patents
A kind of apartment three-dimension object layout generation method based on probabilistic model Download PDFInfo
- Publication number
- CN108470107A CN108470107A CN201810262664.3A CN201810262664A CN108470107A CN 108470107 A CN108470107 A CN 108470107A CN 201810262664 A CN201810262664 A CN 201810262664A CN 108470107 A CN108470107 A CN 108470107A
- Authority
- CN
- China
- Prior art keywords
- room
- model
- furniture
- layout
- pattern
- 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- Withdrawn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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T19/00—Manipulating 3D models or images for computer graphics
- G06T19/20—Editing of 3D images, e.g. changing shapes or colours, aligning objects or positioning parts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2210/00—Indexing scheme for image generation or computer graphics
- G06T2210/04—Architectural design, interior design
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Geometry (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Architecture (AREA)
- Software Systems (AREA)
- Computer Graphics (AREA)
- Civil Engineering (AREA)
- Structural Engineering (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Evolutionary Computation (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of apartment three-dimension object layout generation method based on probabilistic model proposed in the present invention, main contents include：Room modeling, layout generation process, pattern acquiring, layout reasoning, its process is, first from each item data of data concentrated collection room layout, then the Directed Graph Model of a room layout is established, and generate room layout, common pattern is obtained from layout followed by Di Li Cray process mixed models, and new layout more true to nature is finally gone out using the support relation inference between the pattern and object got.The present invention solves the problems, such as that entire room layout can not be generated in the past and object correlation is weak, spatial relationship that can be abundant between captures object, including support relationship and the pattern often occurred, moreover it is possible to generate the layout in entire room and infer new reliable room layout true to nature.
Description
Technical field
The present invention relates to computer vision fields, more particularly, to a kind of apartment three-dimension object cloth based on probabilistic model
Office's generation method.
Background technology
The visual task of indoor scene has been a great concern in recent years, object detection, semanteme including 2D/3D
With example segmentation, room layout estimation, three-dimensional semantic segmentation, depth prediction and positioning immediately and map structuring (SLAM), and
The key of all these tasks is by the entire scene 3D geometries of reasoning and layout, to have better reason to scene
Solution, and considerably reduce the ambiguity of 2D visual signals.The main application that the layout of apartment three-dimension object generates is will be a large amount of
Apartment topology data be converted into computer data, so as to apartment room layout and indoor article counted and located
Reason；Intelligent apartment design can also be carried out, the layout custom by learning mankind designer designs more livable apartment type；
In addition, the method for layout reasoning can also be increased by generating the generated data for including human designer's knowledge for the side of discriminating
The training set of method.However, previous space layout generation method can not all generate the layout in entire room, object can not be also captured
Correlation between body.
The present invention proposes a kind of apartment three-dimension object layout generation method based on probabilistic model, is first adopted from data set
Each item data for collecting room layout, then establishes the Directed Graph Model of a room layout, and generate room layout, followed by
Di Li Cray process mixed models obtain common pattern from layout, finally utilize the branch between the pattern and object got
It holds relation inference and goes out new layout more true to nature.The present invention solves can not generate entire room layout in the past and object is mutual
The problem of relationship weakness, spatial relationship that can be abundant between captures object, including support relationship and the pattern often occurred, also
The layout in entire room can be generated and infer new reliable room layout true to nature.
Invention content
For the problem that can not generate entire room layout and object correlation weakness in the past, the purpose of the present invention exists
In providing a kind of apartment three-dimension object layout generation method based on probabilistic model, first from each of data concentrated collection room layout
Item data, then establishes the Directed Graph Model of a room layout, and generates room layout, mixed followed by Di Li Cray processes
Molding type obtains common pattern from layout, is finally gone out newly using the support relation inference between the pattern and object got
Layout more true to nature.
To solve the above problems, the present invention provides a kind of apartment three-dimension object layout generation method based on probabilistic model,
Its main contents includes：
(1) room models；
(2) layout generation process；
(3) pattern acquiring；
(4) it is laid out reasoning.
Wherein, the described room modeling is and then to parse room by discussing to an apartment topology data collection
It constitutes, establishes a Directed Graph Model, the generating process then defined with model describes this Directed Graph Model, wherein referring to
Apartment topology data collection include 45000 mankind design apartment layout, each layout is crossed by manual verification to be reasonable, and
And there are 8.1 rooms in average often set apartment, most of marks for having room type, such as kitchen or bathroom；These apartments are to use
What one online tool designed, each apartment includes 179 object class, and there are about 2500 CADs in total
(CAD) model and 4,500,000 object instances, each object instance include one with reference to CAD model and it in room
In position and rotation angle.
Further, the composition in the parsing room, is the data set according to reference, is modeled to room, model
Object with different room type (such as kitchen, bathroom, parlor) and corresponding room-size and the inside；Each object
It is parameterized, and is placed in room by the type of CAD model, position and rotation angle；Then intuitively assume in room
Object is all placed in a series of level, and is divided into five classifications：Furniture is located on floor (such as desk, bed) and can support
The big object of other objects；Small object is usually put in the object (such as books or laptop) on furniture；Wall object is
It hangs over or by object (such as picture, shelf) on the wall；Ceiling object is the object (such as lamp) hung on the ceiling；Carpet one
As be to be laid on floor, may there is furniture to be placed on object above；Finally each object of reference data set is assigned to manually
One of class, and in a model separately handle them.
Wherein, the layout generation process is first to be sampled to entire room, then uses complete reference data
Collection is trained, and in the condition distribution that model defines, learns all parameters named by Greek alphabet, in all cases,
Conditional-variable has known value, therefore the parameter of each condition can independently be learnt by Maximum-likelihood estimation, these are all
Directly calculated with closed form.
Further, described that entire room is sampled, it is first to be sampled to room type, removal is uncommon
After possibility, be left the reference data of 22 room types, although room be typically it is several combine, such as kitchen
And dining room, but the possible combination of practical only very small part occurs (22242 kinds are only accounted in kind is possible), so simply using
One classification is distributed to describe whole combination t：
T~Categorical (τ) (1)
Then furniture is sampled, first object sum N is sampled, be respectively then pair under conditions of room type
The classification c of N as n=1 ...nIt is sampled：
N | t~Poisson (λt) (2)
Under different room types, distributed number model N is established for each object class cC, with a small sample (most 4
A example) category distribution and the mixture of Poisson distribution of a large sample indicate this model：
It is symbol for the sake of simplicity, being write as N=∑scNc, regardless of type use n as the index of all object sets, c be used in combinationn
To indicate the type of n-th of object；
In order to make object that overlapping spatially not occur, to the position of furniture object after the definite size for knowing object
It sets and is sampled with room-size, first by each object map a to unit, which is an area of space in room, most
It is just unknown size, and is arranged in a regular structure, it there are free dimension d=(di), can stretch with
Different objects is accommodated, after determining object's position and room-size, the value of these dimensions will be sampled；According to each right
The classification of elephant determines whether each object instance should place (e against walln=1) or it is located in the inside in room, such as
Fruit, which is placed against the wall, sets, then arbitrarily selects a wall
Then, for each face wall, the object order near it and it is assigned to corresponding unit；For distributing to room
Between internal object, select a sequence again, but object is distributed into a cell, use knTo indicate n-th of object
The boundary being assigned to or internal element lattice；
Next, selecting a CAD model and its rotation angle for each object instance；They are determined together
The size in the space occupied by it, CAD model mnBe under conditions of object class according to it is in the corner or edge in room, also
It is that internal independence selects：
It is assumed that rotation θnOnly with respect to vertical axis, and with based on cnOne using pi/2 as multiple discrete uniform distribution and
The mixed distribution of one continuous uniform distribution between [0,2 π] to model to it；
To the modeling of ceiling object with to furniture, first being adopted to the quantity and rotation angle of each related object class
Sample, wherein only being sampled once to the CAD model of each object class；For space cell lattice structure, then a simple net is used
Lattice when selecting the size in room, then need to consider required space if it is more than the required space of furniture；
The processing mode of carpet is identical with the object on ceiling, but it only takes up an individual layer, and CAD
Model samples each example primary；
Then the precise physical location to selected objects above and final room-size sample so that all pairs
As all non-intersecting, a CAD model m is being established for each objectnWith rotation angle θnLater, so that it may with solid with an axis
Fixed bounding box calculates its spatial dimension, it is contemplated that the k of spatial dimension and unit distributionn, setting unit structure can be passed through
Each free dimension di, keep its as small as possible, calculates the minimum room-size that all objects of sening as an envoy to are suitable for, in practice,
This can make room excessively crowded, so sampling some fillings p in each data collection with a four-dimensional diagonal Gaussian Profilen：
Wherein Normal+It is the isotropism Gaussian Profile by positive limiting, then again free dimension diIt is set as most
It is small, obtain minimum room-size；
And then the small object on each furniture is sampled, for every a piece of furniture n, to the object class being placed on above itQuantityIt is sampled, is similarly to the sampling to article of furniture quantity, but be the type of the furniture as support now
cn：
Wherein positionIt is set as with supporting the region above furniture, CAD model mnWith rotation angle θnAlso with family
Tool is the same,Index as object on all furniture.
Wherein, the pattern acquiring is to be indicated by Mode exploration, pattern, carry out three steps of sampling to schema instance
Suddenly, some patterns are added in a model, i.e., certain specifically spatial relationships more more conventional than other relationships between object.
Further, the Mode exploration is all training sets searched for associated class tuple and occurred, i.e., from a room
Then one element assignment of tuple is base object by the middle all arrangements for obtaining the K object instances for belonging to associated class, and right
In each event, the relative displacement from base object center to other object centers is calculated, these displacements can be regarded as 2 (K-
1) point in dimension space；Using variable reasoning fitting Di Li Cray process mixed models and Gauss cluster with diagonal covariance into
Row cluster；Generate each cluster be a candidate pattern, and finally according to it whether with enough dimensional compactness come
Determine that it is a pattern, i.e. the relative position of schema elements can specifically be calculated 2 σ wheels by constraint stringent enough
The position of each element in region in exterior feature in pattern, if as soon as all these both less than threshold values, then this cluster
Regard as a pattern.
Further, the pattern indicates, is indicated with the nonparametric of a pattern：In training set, directly store
Relative position, relative direction and CAD model.
Further, described that schema instance is sampled, it is that given pattern is needed to merge them into life
At in the process, using schema instance as complete unit, it is similar to singleton object, more precisely, by each type of pattern
Identical processing is carried out with object class, and the example quantity based on room type is sampled；It then will each complete mould
Formula example allocation selects a storage training set example to define each example to the individual unit lattice in room layout
Relative position/direction of the pairs of elephant of modal sets and CAD model；Then, it is converted all composition objects as a unit
And rotation, schema instance is put in a room；
And this requires learn the radix of each pattern distribution (point of the instance number in i.e. each room based on room type
Cloth), it is also necessary to correspondingly adjust existing class radix parameter πtc,Show now with less singleton.
Wherein, the layout reasoning is weighed to the content in each room using model according to reference data set
New sampling, but the existing structure for being to maintain apartment is constant, more precisely, the size based on room and type, to furniture and small
Object is sampled, including the object on the object and ceiling on wall；
Due to the conditional sampling structure of model, the adjusting to room type and wall/ceiling object is very simple；So
And it is much more complex to the adjusting of room-size, because of this class and size dependent on furniture object in different-style；Therefore, make
With approximate method collecting sample from qualified result, specifically, the type just for room and wall/day
Sample is generated on card, rather than it is bigger than desired size more than 5 centimetres finally to exclude those bulks for room-size
Sample.
Description of the drawings
Fig. 1 is a kind of system flow chart of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.
Fig. 2 is a kind of generation illustraton of model of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.
Fig. 3 is that a kind of units of furniture of the apartment three-dimension object layout generation method based on probabilistic model of the present invention divides
Figure.
Fig. 4 is a kind of small object sample graph of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.
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 chart of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.It is main
To include that room models, layout generates, pattern acquiring, is laid out reasoning.
Room modeling be by discussing to an apartment topology data collection, and then parse room composition, establish one
A Directed Graph Model, the generating process then defined with model describe this Directed Graph Model, wherein the apartment layout referred to
Data set includes the apartment layout of 45000 mankind's designs, and each layout is crossed by manual verification to be reasonable, and is averagely often covered
There are 8.1 rooms in apartment, most of marks for having room type, such as kitchen or bathroom；These apartments are with an online work
Tool design, each apartment includes 179 object class, in total there are about 2500 CAD (CAD) models and
4500000 object instances, each object instance include one with reference to CAD model and its position and rotation in a room
Angle.
Further, the composition in the parsing room, is the data set according to reference, is modeled to room, model
Object with different room type (such as kitchen, bathroom, parlor) and corresponding room-size and the inside；Each object
It is parameterized, and is placed in room by the type of CAD model, position and rotation angle；Then intuitively assume in room
Object is all placed in a series of level, and is divided into five classifications：Furniture is located on floor (such as desk, bed) and can support
The big object of other objects；Small object is usually put in the object (such as books or laptop) on furniture；Wall object is
It hangs over or by object (such as picture, shelf) on the wall；Ceiling object is the object (such as lamp) hung on the ceiling；Carpet one
As be to be laid on floor, may there is furniture to be placed on object above；Finally each object of reference data set is assigned to manually
One of class, and in a model separately handle them.
And pattern acquiring is to be indicated by Mode exploration, pattern, carry out three steps of sampling to schema instance, in a model
Some patterns are added, i.e., certain specifically spatial relationships more more conventional than other relationships between object.
Further, the Mode exploration is all training sets searched for associated class tuple and occurred, i.e., from a room
Then one element assignment of tuple is base object by the middle all arrangements for obtaining the K object instances for belonging to associated class, and right
In each event, the relative displacement from base object center to other object centers is calculated, these displacements can be regarded as 2 (K-
1) point in dimension space；Using variable reasoning fitting Di Li Cray process mixed models and Gauss cluster with diagonal covariance into
Row cluster；Generate each cluster be a candidate pattern, and finally according to it whether with enough dimensional compactness come
Determine that it is a pattern, i.e. the relative position of schema elements can specifically be calculated 2 σ wheels by constraint stringent enough
The position of each element in region in exterior feature in pattern, if as soon as all these both less than threshold values, then this cluster
Regard as a pattern.
Further, the pattern indicates, is indicated with the nonparametric of a pattern：In training set, directly store
Relative position, relative direction and CAD model.
Further, described that schema instance is sampled, it is that given pattern is needed to merge them into life
At in the process, using schema instance as complete unit, it is similar to singleton object, more precisely, by each type of pattern
Identical processing is carried out with object class, and the example quantity based on room type is sampled；It then will each complete mould
Formula example allocation selects a storage training set example to define each example to the individual unit lattice in room layout
Relative position/direction of the pairs of elephant of modal sets and CAD model；Then, it is converted all composition objects as a unit
And rotation, schema instance is put in a room；
And this requires learn the radix of each pattern distribution (point of the instance number in i.e. each room based on room type
Cloth), it is also necessary to correspondingly adjust existing class radix parameter πtc,Show now with less singleton.
It is to have carried out resampling to the content in each room using model, still according to reference data set to be laid out reasoning
Keep the existing structure in apartment constant, more precisely, the size based on room and type take furniture and small articles
The object on object and ceiling on sample, including wall；
Due to the conditional sampling structure of model, the adjusting to room type and wall/ceiling object is very simple；So
And it is much more complex to the adjusting of room-size, because of this class and size dependent on furniture object in different-style；Therefore, make
With approximate method collecting sample from qualified result, specifically, the type just for room and wall/day
Sample is generated on card, rather than it is bigger than desired size more than 5 centimetres finally to exclude those bulks for room-size
Sample.
Fig. 2 is a kind of generation illustraton of model of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.Cloth
The detailed process that office generates model is first to be sampled to entire room, is then trained using complete reference data set,
In the condition distribution that model defines, learn all parameters named by Greek alphabet, in all cases, conditional-variable has
Known value, therefore the parameter of each condition can independently be learnt by Maximum-likelihood estimation, these are all direct closings
Form calculate.
It is first to be sampled to room type wherein to carry out sampling to entire room, after removing uncommon possibility,
Be left the reference data of 22 room types, although room be typically it is several combine, such as kitchen and dining room are real
Only have the possible combination of very small part to occur (2 in border2242 kinds are only accounted in kind is possible), so being simply distributed with a classification
To describe whole combination t：
T~Categorical (τ) (1)
Then furniture is sampled, first object sum N is sampled, be respectively then pair under conditions of room type
The classification c of N as n=1 ...nIt is sampled：
N | t~Poisson (λt) (2)
Under different room types, distributed number model N is established for each object class cC, with a small sample (most 4
A example) category distribution and the mixture of Poisson distribution of a large sample indicate this model：
It is symbol for the sake of simplicity, being write as N=∑scNc, regardless of type use n as the index of all object sets, c be used in combinationn
To indicate the type of n-th of object；
After dividing the corresponding unit of good each object instance, for each object instance select a CAD model and it
Rotation angle；They determine the size in the space occupied by it, CAD model m togethernIt is under conditions of object class according to it
It is to be selected in the corner in room or edge or internal independence：
It is assumed that rotation θnOnly with respect to vertical axis, and with based on cnOne using pi/2 as multiple discrete uniform distribution and
The mixed distribution of one continuous uniform distribution between [0,2 π] to model to it；
To the modeling of ceiling object with to furniture, first being adopted to the quantity and rotation angle of each related object class
Sample, wherein only being sampled once to the CAD model of each object class；For space cell lattice structure, then a simple net is used
Lattice when selecting the size in room, then need to consider required space if it is more than the required space of furniture；
The processing mode of carpet is identical with the object on ceiling, but it only takes up an individual layer, and CAD
Model samples each example primary；
Then the precise physical location to selected objects above and final room-size sample so that all pairs
As all non-intersecting, a CAD model m is being established for each objectnWith rotation angle θnLater, so that it may with solid with an axis
Fixed bounding box calculates its spatial dimension, it is contemplated that the k of spatial dimension and unit distributionn, setting unit structure can be passed through
Each free dimension di, keep its as small as possible, calculates the minimum room-size that all objects of sening as an envoy to are suitable for, in practice,
This can make room excessively crowded, so sampling some fillings p in each data collection with a four-dimensional diagonal Gaussian Profilen：
Wherein Normal+It is the isotropism Gaussian Profile by positive limiting, then again free dimension diIt is set as most
It is small, obtain minimum room-size；
Fig. 3 is that a kind of units of furniture of the apartment three-dimension object layout generation method based on probabilistic model of the present invention divides
Figure.So that make object that overlapping spatially not occur, after the definite size for knowing object to the position of furniture object and
Room-size is sampled, and first by each object map a to unit, which is an area of space in room, is initially
Unknown size, and be arranged in a regular structure, it has free dimension d=(di), it can stretch to accommodate
Different objects will sample the value of these dimensions after determining object's position and room-size；According to each object
Classification determines whether each object instance should place (e against walln=1) or it is located in the inside in room, if leaned on
Wall is placed, then arbitrarily selects a wall
Then, for each face wall, the object order near it and it is assigned to corresponding unit；For distributing to room
Between internal object, select a sequence again, but object is distributed into a cell, use knTo indicate n-th of object
The boundary being assigned to or internal element lattice；
Fig. 4 is a kind of small object sample graph of the apartment three-dimension object layout generation method based on probabilistic model of the present invention.
To the object class being placed on every a piece of furniture nQuantityIt is sampled, is similarly to adopt article of furniture quantity
Sample, but be the type c of the furniture as support nown：
Wherein positionIt is set as with supporting the region above furniture, CAD model mnWith rotation angle θnAlso with family
Tool is the same,Index as object on all furniture.
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 apartment three-dimension object based on probabilistic model is laid out generation method, which is characterized in that include mainly that room models
(1) layout generation process (two)；Pattern acquiring (three)；It is laid out reasoning (four).
2. based on described in claims 1 room model (one), which is characterized in that by an apartment topology data collection into
Row discusses, and then parses the composition in room, establish a Directed Graph Model, and the generating process then defined with model describes this
A Directed Graph Model is each laid out quilt wherein the apartment topology data collection referred to includes the apartment layout of 45000 mankind's designs
Manual verification, which crosses, to be reasonable, and there are 8.1 rooms in average often set apartment, most of marks for having room type, such as kitchen
Room or bathroom；These apartments are designed with an online tool, each apartment include 179 object class, in total there are about
2500 CAD (CAD) models and 4,500,000 object instances, each object instance include a reference
CAD model and its position and rotation angle in a room.
3. the composition based on the parsing room described in claims 2, which is characterized in that according to the data set of reference, to room
It is modeled, model has the object of different room types (such as kitchen, bathroom, parlor) and corresponding room-size and the inside
Body；Each object is parameterized by the type of CAD model, position and rotation angle, and is placed in room；Then intuitively
Assuming that the object in room is all placed in a series of level, and it is divided into five classifications：Furniture is located on floor (such as table
Son, bed) it can support the big objects of other objects；Small object is usually put in object (such as books or notebook electricity on furniture
Brain)；Wall object is hung over or by object (such as picture, shelf) on the wall；Ceiling object is the object hung on the ceiling
Body (such as lamp)；Carpet is usually to be laid on floor, may have furniture to be placed on object above；Finally manually by reference data set
Each object assign to one of class, and they are separately handled in a model.
4. based on the layout generation process (two) described in claims 1, which is characterized in that first entire room is sampled,
Then it is trained using complete reference data set, in the condition distribution that model defines, what study was named by Greek alphabet
All parameters, in all cases, conditional-variable have known value, therefore the parameter of each condition can pass through maximum likelihood
Estimation independently learns, these are all directly calculated with closed form.
5. based on being sampled to entire room described in claims 4, which is characterized in that first adopted to room type
Sample after removing uncommon possibility, is left the reference data of 22 room types, exists although room is typically several combinations
Together, such as kitchen and dining room, but practical only have the possible combination generation of very small part (22242 are only accounted in kind is possible
Kind), so simply describing whole combination t with a classification distribution：
T~Categorical (τ) (1)
Then furniture is sampled, first object sum N is sampled, be respectively then object n under conditions of room type
The classification c of=1 ... NnIt is sampled：
N | t~Poisson (λt) (2)
Under different room types, distributed number model N is established for each object class cC, with a small sample (most 4 realities
Example) category distribution and the mixture of Poisson distribution of a large sample indicate this model：
It is symbol for the sake of simplicity, being write as N=∑scNc, regardless of type use n as the index of all object sets, c be used in combinationnCarry out table
Show the type of n-th of object；
So that make object that overlapping spatially not occur, after the definite size for knowing object to the position of furniture object and
Room-size is sampled, and first by each object map a to unit, which is an area of space in room, is initially
Unknown size, and be arranged in a regular structure, it has free dimension d=(di), it can stretch to accommodate
Different objects will sample the value of these dimensions after determining object's position and room-size；According to each object
Classification determines whether each object instance should place (e against walln=1) or it is located in the inside in room, if leaned on
Wall is placed, then arbitrarily selects a wall
Then, for each face wall, the object order near it and it is assigned to corresponding unit；For distributing in room
The object in portion selects a sequence, but object is distributed to a cell again, uses knTo indicate n-th of object distribution
The boundary arrived or internal element lattice；
Next, selecting a CAD model and its rotation angle for each object instance；They determine its institute together
The size in the space occupied, CAD model mnBe under conditions of object class according to it is at the corner in room or edge or interior
Portion's independent choice：
It is assumed that rotation θnOnly with respect to vertical axis, and with based on cnOne using pi/2 as multiple discrete uniform distribution and one
The mixed distribution of continuous uniform distribution between [0,2 π] to model to it；
To the modeling of ceiling object with to furniture, the first quantity to each related object class and rotation angle sampling,
In the CAD model of each object class is only sampled once；For space cell lattice structure, then a simple grid is used, such as
It is more than the required space of furniture to fruit, when selecting the size in room, then needs to consider required space；
The processing mode of carpet is identical with the object on ceiling, but it only takes up an individual layer, and CAD model
It is primary to the sampling of each example；
Then the precise physical location to selected objects above and final room-size sample so that all objects are all
It is non-intersecting, establishing a CAD model m for each objectnWith rotation angle θnLater, so that it may with fixed with an axis
Bounding box calculates its spatial dimension, it is contemplated that the k of spatial dimension and unit distributionn, each of setting unit structure can be passed through
Free dimension di, keep its as small as possible, calculate the minimum room-size that all objects of sening as an envoy to are suitable for, in practice, this meeting
Keep room excessively crowded, so sampling some fillings p in each data collection with a four-dimensional diagonal Gaussian Profilen：
Wherein Normal+It is the isotropism Gaussian Profile by positive limiting, then again free dimension diIt is set as minimum, obtains
To minimum room-size；
And then the small object on each furniture is sampled, for every a piece of furniture n, to the object class being placed on above it's
QuantityIt is sampled, is similarly to the sampling to article of furniture quantity, but be the type c of the furniture as support nown：
Wherein positionIt is set as with supporting the region above furniture, CAD model mnWith rotation angle θnAlso with furniture one
Sample,Index as object on all furniture.
6. based on the pattern acquiring (three) described in claims 1, which is characterized in that indicated, to mould by Mode exploration, pattern
Formula example carries out three steps of sampling, adds some patterns in a model, i.e., certain specifically than other relationships between object
More conventional spatial relationship.
7. based on the Mode exploration described in claims 6, which is characterized in that all training that search associated class tuple occurs
Collection obtains all arrangements for the K object instances for belonging to associated class, then by an element assignment of tuple that is, from a room
For base object, and for each event, calculate the relative displacement from base object center to other object centers, these displacements
It can be regarded as the point in 2 (K-1) dimension spaces；Using the Di Li Cray process mixed models of variable reasoning fitting and with diagonal association
The Gauss cluster of variance clusters；Whether each cluster generated is a candidate pattern, and finally had according to it enough
Dimensional compactness determine that it is a pattern, i.e., the relative position of schema elements is by constraint stringent enough, tool
Body can calculate the position of each element in pattern in the region in 2 σ profiles, if an all these both less than threshold value,
So this cluster is it is assumed that be a pattern.
8. being indicated based on the pattern described in claims 6, which is characterized in that indicated using the nonparametric of a pattern：It is instructing
Practice and concentrate, directly stores relative position, relative direction and CAD model.
9. based on being sampled to schema instance described in claims 6, which is characterized in that for given pattern, need
It merges them into generating process, using schema instance as complete unit, is similar to singleton object, it more precisely, will
Each type of pattern carries out identical processing with object class, and is sampled to the example quantity based on room type；Then
Each complete schema instance is distributed into the individual unit lattice in room layout, for each example, selects a storage instruction
Practice relative position/direction and CAD model that collection example carrys out defining mode composition object；Then, using all composition objects as one
A unit is converted and is rotated, and schema instance is put in a room；
And this requires learning the radix of each pattern based on room type to be distributed (distribution of the instance number in i.e. each room), also
It needs correspondingly to adjust existing class radix parameter πtc,Show now with less singleton.
10. based on the layout reasoning (four) described in claims 1, which is characterized in that according to reference data set, use model pair
The content in each room has carried out resampling, but the existing structure for being to maintain apartment is constant, more precisely, based on room
Size and type, are sampled furniture and small articles, including the object on the object and ceiling on wall；
Due to the conditional sampling structure of model, the adjusting to room type and wall/ceiling object is very simple；However,
It is much more complex to the adjusting of room-size, because of this class and size dependent on furniture object in different-style；Therefore, using one
A approximate method collecting sample from qualified result, specifically, the type just for room and wall/ceiling
Upper generation sample, rather than room-size finally exclude those bulks than the big sample more than 5 centimetres of desired size.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810262664.3A CN108470107A (en) | 2018-03-28 | 2018-03-28 | A kind of apartment three-dimension object layout generation method based on probabilistic model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810262664.3A CN108470107A (en) | 2018-03-28 | 2018-03-28 | A kind of apartment three-dimension object layout generation method based on probabilistic model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108470107A true CN108470107A (en) | 2018-08-31 |
Family
ID=63265933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810262664.3A Withdrawn CN108470107A (en) | 2018-03-28 | 2018-03-28 | A kind of apartment three-dimension object layout generation method based on probabilistic model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108470107A (en) |
-
2018
- 2018-03-28 CN CN201810262664.3A patent/CN108470107A/en not_active Withdrawn
Non-Patent Citations (1)
| Title |
|---|
| PAUL HENDERSON等: "A Generative Model of 3D Object Layouts in Apartments", 《HTTPS://ARXIV.ORG/ABS/1711.10939V1》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109871604B (en) | Indoor function division method based on depth countermeasure network model | |
| WO2021008566A1 (en) | Artificial intelligence systems and methods for interior design | |
| US20200242849A1 (en) | Methods and systems for rendering and modifying three-dimensional models for interior design | |
| US10049297B1 (en) | Data driven method for transferring indoor scene layout and color style | |
| EP4026037A1 (en) | Systems and methods for structural design using modeling and simulation for architectural planning | |
| WO2020079494A1 (en) | 3d scene synthesis techniques using neural network architectures | |
| CN109635843A (en) | A kind of three-dimensional object model classification method based on multi-view image | |
| CN110378997A (en) | A kind of dynamic scene based on ORB-SLAM2 builds figure and localization method | |
| JP2010262646A (en) | Method and apparatus of animation planning for dynamic graph | |
| Chen et al. | Extraction of indoor objects based on the exponential function density clustering model | |
| CN107240151A (en) | A kind of scene layout based on parent-child constraint preserves and reproducting method | |
| CN111091618B (en) | Combined indoor layout method and system based on house type data driving | |
| CN103198522A (en) | Three-dimensional scene model generation method | |
| CN107045519B (en) | A kind of visualization system of analysis crowd movement law | |
| WO2023124160A1 (en) | Method, system and apparatus for automatically generating three-dimensional house layout, and medium | |
| CN105957148B (en) | A kind of granularity equalization data method for organizing of complex three-dimensional building model | |
| Zhao et al. | DAR-Net: Dynamic aggregation network for semantic scene segmentation | |
| US20220358739A1 (en) | Method, device and computer program product for manipulating virtual bounding volumes | |
| CN108921952A (en) | Object functionality prediction technique, device, computer equipment and storage medium | |
| CN105809742B (en) | A kind of indoor scene color method of data-driven | |
| CN108470107A (en) | A kind of apartment three-dimension object layout generation method based on probabilistic model | |
| CN113946900A (en) | Method for quickly recommending similar house types based on house type profiles and distribution characteristics | |
| Yu et al. | 3DSAC: Size Adaptive Clustering for 3D object detection in point clouds | |
| KR102492699B1 (en) | Apparatus and method for recommending object on adjacent trunk line based on data structure of trunk line connection generated from big data of interior service | |
| WO2023174555A1 (en) | Computer-implemented reconstruction of interior rooms |
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: 20180831 |
|
| WW01 | Invention patent application withdrawn after publication |