Calculation method of three-dimensional model's spherical expression based on multi-stage deformation reconstruction
The embodiments of the present disclosure disclose a three-dimensional model's spherical expression calculation method. A specific implementation of the method includes: processing an input three-dimensional model into a dense point cloud model; inputting the dense point cloud model to a multi-level feature extraction module, extracting high-dimensional feature vectors of different abstract levels; inputting the high-dimensional feature vectors, template ball point cloud and abstract preprocessed point cloud model to a point cloud deformation reconstruction module to obtain a deformed reconstruction point cloud model; extracting multi-stage deformation reconstruction process information, combining the multi-stage deformation reconstruction process information with the template ball point cloud to together form complete information describing the three-dimensional model; obtaining a density correspondence from the three-dimensional model to the template ball point cloud and a density correspondence between different three-dimensional models. This implementation does not require time-consuming and labor-intensive manual annotations, improving the efficiency of characterization learning.
1. A method for ball expression calculation for a three-dimensional model, comprising:
step 1: processing an input three-dimensional model into a dense point cloud model, and performing multi-resolution point cloud abstract preprocessing, to obtain an abstract preprocessed point cloud model;
step 2: inputting the dense point cloud model into a feature extraction module, down-sampling the point cloud for a predetermined number of times and extracting high-dimensional feature vectors of different abstract levels, to obtain a set of high-dimensional feature vectors;
step 3: inputting the high-dimensional feature vectors, a template ball point cloud and the abstract preprocessed point cloud model to a point cloud deformation reconstruction module to generate a reconstructed point cloud model, and repeating step 3 for a predetermined number of times, to obtain a deformed reconstruction point cloud model;
step 4: extracting multi-stage deformation reconstruction process information in step 3, combining the multi-stage deformation reconstruction process information with the template ball point cloud to together form complete information describing the three-dimensional model to obtain a three-dimensional model's spherical expression, wherein the multi-stage deformation reconstruction process information includes a deformation reconstruction point-by-point offset and a high-dimensional feature vector;
step 5: based on the three-dimensional model's spherical expression, obtaining a density correspondence from the three-dimensional model to the template ball point cloud and a density correspondence between different three-dimensional models, wherein the density correspondence is suitable for a variety of three-dimensional geometric analysis and editing applications.
2. The method of claim 1 , wherein, the processing an input three-dimensional model into a dense point cloud model and performing multi-resolution point cloud abstract preprocessing to obtain an abstract preprocessed point cloud model, comprising:
step 1.1: performing a point cloud down-sampling process to the three-dimensional model, to generate relatively sparse point cloud models;
step 1.2: performing Gaussian redistribution to the relatively sparse point cloud models, to generate the abstract preprocessed point cloud model.
3. The method of claim 2 , wherein, the inputting the high-dimensional feature vectors, template ball point cloud and the abstract preprocessed point cloud model to a point cloud deformation reconstruction module to generate a reconstructed point cloud model and repeating step 3 for a predetermined number of times to obtain a deformed reconstruction point cloud model, comprising:
step 3.1: inputting the abstract preprocessed point cloud model into the point cloud deformation reconstruction module;
step 3.2: based on the template ball point cloud and the high-dimensional feature vector, predicting the point-by-point deformation offset through a deep neural network, then outputting the reconstruction point cloud model after deformation, to obtain a deformed reconstruction point cloud model, wherein, a deformation reconstruction optimizing process includes two constraint target conditions.
4. The method of claim 3 , wherein, the constraint target conditions include a multi-stage shape similarity constraint condition and a point set deformation topological consistency constraint condition, the multi-stage shape similarity constraint condition is a constraint to, in a multi-stage deformation reconstruction process, keep reconstruction results of each stage and the preprocessed multi-resolution point cloud abstract preprocessing result, while the point set deformation topological consistency constraint condition is a constraint to keep deformation offset of adjacent points in step 3 similar in the multi-stage deformation reconstruction process, as a deformation regular term constraint, and the multi-stage deformation reconstruction process comprises constraint target conditions.
5. The method of claim 4 , wherein, for any deformation stage, a chamfer distance CD is used to measure a similarity between a reconstructed point cloud and a target point cloud, a point set distance is CD, and a shape similarity constraint condition is:
L
C
D
(
k
)
=
∑
p
i
∈
S
k
min
q
j
∈
T
k
p
i
-
q
j
2
2
+
∑
q
i
∈
T
k
min
p
j
∈
S
k
q
j
-
p
i
2
2
,
wherein, k represents a serial number of the abstract levels, L CD (k) represents a CD of a k th deformation stage S k and T k , S k represents the abstract preprocessed point cloud model of a k th abstract level, T k is the deformation reconstruction model of the k th abstract level, p i represents a point in the point cloud model S k , q j represents a point in the point cloud model T k ,
above formula calculates a distance between two point sets, a specific calculation method being to, for any point in set S k , calculate square of a distance between arbitrary point and a nearest point of the arbitrary point in set T k , this measure is symmetrical;
the multi-stage shape similarity constraint condition is:
L 1 ( k )=Σ k=K K L CD ( k ),
wherein, L 1 (k) represents a sum of the CD of the abstract preprocessed point cloud model and the deformed reconstruction point cloud model in a k th deformation stage, k represents the serial number of the abstract levels, K represents a number of the abstract levels, and L CD (k) represents the CD of the k th deformation stage i.e. S k and T k .
6. The method of claim 5 , wherein, in the multi-stage deformation reconstruction process, the deformation offsets of the adjacent points are consistent, and for any adjacent deformation stage, a point-to-point Euclidean distance in the template is first calculated, thereby constructing a penalty coefficient ω i,j , meeting the farther between two points i and j, the smaller the penalty coefficient ω i,j , let a predicted offset of a point of the k level be d k (i), then the point set deformation topological consistency constraint condition formula is:
ω i,j =exp(−∥ p i −p j ∥ 2 ),∀ p i ∈T k ,∀p j ∈T k ,
wherein, i represents serial number, j represents serial number, ω i,j represents the penalty coefficient, p i represents a i th point in T k , P j represents a j th point in T k ;
L reg ( k )=Σ 0<i≠j≤N ω i,j ∥d k ( i )− d k ( j )∥ 2 ,
wherein, k represents the serial number of the abstract levels, L reg (k) represents the regular term constraint of the k th abstract level, ω i,j represents the penalty coefficient, i represents the serial number, j represents the serial number, d k (i) represents the predicted offset of the i th point in T k , d k (j) represents the predicted offset of the j th point in T k ;
the multi-stage point set deformation topological consistency constraint condition is:
L 2 ( k )=Σ k=K K L reg ( k ),
wherein, L 2 (k) represents the regular term constraint of the multi-stage abstract level, k represents serial number of the abstract levels, K represents a number of the abstract levels, and L reg (k) represents the regular term constraint of the k th abstract level.
7. The method of claim 6 , wherein, the three-dimensional model's spherical expression is:
CS:={T K D k ,z k |K =1, . . . , K},
wherein, CS: represents the three-dimensional model's spherical expression, T K represents the deformation reconstruction model of the k th abstract level, and is also the template ball point cloud, D k represents the point-by-point offset of the k th abstract level, z k represents the high-dimensional feature vector of the k th abstract level, k represents the serial number of the abstract levels, and K represents the number of the abstract levels.