Depth estimation method for monocular image based on multi-scale CNN and continuous CRF
A depth estimation method for a monocular image based on a multi-scale CNN and a continuous CRF is disclosed in this invention. A CRF module is adopted to calculate a unary potential energy according to the output depth map of a DCNN, and the pairwise sparse potential energy according to input RGB images. MAP (maximum a posteriori estimation) algorithm is used to infer the optimized depth map at last. The present invention integrates optimization theories of the multi-scale CNN with that of the continuous CRF. High accuracy and a clear contour are both achieved in the estimated depth map; the depth estimated by the present invention has a high resolution and detailed contour information can be kept for all objects in the scene, which provides better visual effects.
1. A depth estimation method for a monocular image based on a multi-scale CNN (convolution neural network) and a continuous CRF (conditional random field), comprising steps of:
(1) obtaining a preset amount of sample images with a camera, and processing the sample images with data enhancement and normalization in sequence with a computer;
(2) training a DCNN (deep convolution neural network) with normalized sample images by two stages with the computer, wherein the DCNN comprises a first stack, a second stack and a third stack; the first stack and the second stack are trained by a first stage, and the third stack is trained by a second stage;
(3) for an input image I whose depth is to be estimated, establishing a conditional probability distribution function P(Y|I) related to the input image I and a depth map Y thereof according to a DCNN output:
P
(
Y
|
I
)
=
1
Z
(
I
)
exp
(
-
E
(
Y
,
I
)
)
Z
(
I
)
=
∫
Y
exp
(
-
E
(
Y
,
I
)
)
dY
E
(
Y
,
I
)
=
∑
i
(
z
i
-
y
i
)
2
+
∑
ij
(
y
i
-
y
j
)
2
[
w
ij
1
exp
(
-
p
i
-
p
j
2
2
σ
ij
1
2
-
c
i
-
c
j
2
2
σ
ij
2
2
)
+
w
ij
2
exp
(
-
p
i
-
p
j
2
2
σ
ij
3
2
)
]
wherein Z(I) is a partition function corresponding to the input image I, E(Y,I) is an energy function related to the input image I and the depth map Y thereof, z i is a predicted logarithmic depth value of a No. i effective pixel, y i and y j are true logarithmic depth values of the No. i effective pixel and a No. j effective pixel respectively, p i and p j are coordinate values of the No. i effective pixel and the No. j effective pixel respectively, c i and c j are LAB color values of the No. i effective pixel and the No. j effective pixel respectively; w ij1 , w ij2 , σ ij1 , σ ij2 and σ ij3 are all CRF model parameters; i and j are both natural numbers and 1≤i≤N, 1≤j≤N, N is a total effective pixel number;
then finding a solution to maximize P(Y|I), and converting a calculated depth map Ŷ from a logarithmic space into a linear space, so as to obtain a depth map with a high accuracy and with object contour detail reserved; and
(3) forming a three-dimensional model of the input image I using the depth map, and projecting the three-dimensional model of the input image I to obtain an image from another visual angle of the input image I, or calculating the image from another visual angle directly with the depth map according to a binocular disparity principle.
2. The depth estimation method, as recited in claim 1 , wherein the data enhancement comprises scaling, rotating, translating, luminance shifting, and flipping; the normalization comprises subtracting a mean image of the sample images and dividing by a mean square error.
3. The depth estimation method, as recited in claim 1 , wherein the first stack of the DCNN is an Imagenet pre-trained network, which adopts first seven layers of an Alexnet module or first fifteen layers of a VGG16 module; the second stack is formed by five convolution layers, first four of the five convolution layers are proceeded with pooling, a last of the five convolution layers has a linear output; wherein an input of a first convolution layer of the five convolution layers is the normalized sample images obtained in the step (1), and an input of a second convolution layer of the five convolution layers is a feature formed by concatenating an up-sampled output of the first stack with a pooling output of the first convolution layer of the five convolution layers; all inputs of a later convolution layers is an output of a previous convolution layer thereof; the third stack is formed by four convolution layers, first three thereof are proceeded with pooling, a last of the four convolution layers has a linear output; wherein the input of a first convolution layer of the four convolution layers is the normalized sample images obtained in the step (1), and an input of a second convolution layer of the four convolution layers is a feature formed by concatenating an up-sampled output of the second stack with a pooling output of the first convolution layer of the four convolution layers; all inputs of a later convolution layer is an output of a previous convolution layer thereof.
4. The depth estimation method, as recited in claim 3 , wherein a loss function L used in the first stage during training is:
L
=
1
N
∑
i
d
i
2
-
1
2
N
2
(
∑
i
d
i
)
2
+
1
N
∑
i
[
(
∇
x
d
i
)
2
+
(
∇
y
d
i
)
2
]
wherein d i =lgx i −lgx i *, x i and x i * are a predicted depth value and a ground truth of the No. i effective pixel respectively, ∇ x d i and ∇ y d i are a horizontal gradient and a vertical gradient of d i respectively.
5. The depth estimation method, as recited in claim 1 , wherein in the step (3), the solution to maximize P(Y|I) is achieved by:
Y
~
=
argmax
Y
P
(
Y
|
I
)
=
A
-
1
Z
A
=
E
+
D
-
S
wherein Z is a depth value after up-sampling and boundary completion an output of the third stack of the DCNN with bilinear interpolation, D is a diagonal matrix with diagonal elements d ii =Σ j s ij , is a similarity matrix whose element at a row i and a column j is
s
ij
=
∑
ij
(
y
i
-
y
j
)
2
[
w
ij
1
exp
(
-
p
i
-
p
j
2
2
σ
ij
1
2
-
c
i
-
c
j
2
2
σ
ij
2
2
)
+
w
ij
2
exp
(
-
p
i
-
p
j
2
2
σ
ij
3
2
)
]
,
E is an N×N unit matrix.
6. The depth estimation method, as recited in claim 1 , wherein the CRF model parameters w ij1 , w ij2 , σ ij1 , σ ij2 and σ ij3 are obtained by: integrating the CRF into the DCNN, and optimizing with an SGD method, wherein loss is calculated by comparing a CRF output with a ground truth logarithmic depth map; or independently optimizing a CRF module by using cross validation, wherein optimized parameters are searched by cycling from large step sizes to small step sizes in a certain range this time the DCNN parameters are fixed.