Disparity estimation method and apparatus, and image processing device and storage medium
Provided are a disparity estimation method and apparatus, an image processing device, and a storage medium. The method includes: processing input images based on a first network model to obtain a direct cost volume of the input images, where the input images include a first image and a second image, and the first network model includes a convolutional neural network, a pyramid convolutional network, and a spatial pyramid pooling layer; processing the input images based on a second network model to obtain an associated cost volume of the input images, where the second network model includes a residual network; determining an estimated cost of the input images according to the associated cost volume and the direct cost volume; and calculating an estimated disparity corresponding to the first image and the second image according to the estimated cost.
1 . A disparity estimation method, comprising:
processing input images based on a first network model to obtain a direct cost volume of the input images, wherein the input images comprise a first image and a second image, and the first network model comprises a convolutional neural network, a pyramid convolutional network, and a spatial pyramid pooling layer;
processing the input images based on a second network model to obtain an associated cost volume of the input images, wherein the second network model comprises a residual network;
determining an estimated cost of the input images according to the associated cost volume and the direct cost volume; and
Calculating an estimated disparity corresponding to the first image and the second image according to the estimated cost.
2 . The method of claim 1 , wherein processing the input images based on the first network model to obtain the direct cost volume of the input images comprises:
extracting feature information of the first image and feature information of the second image separately based on the convolutional neural network;
performing a convolution operation on the feature information of the first image and the feature information of the second image separately based on the pyramid convolutional network to obtain multi-scale feature information of the first image and multi-scale feature information of the second image;
aggregating the multi-scale feature information of the first image and the multi-scale feature information of the second image separately based on the spatial pyramid pooling layer to obtain matching costs of the first image and matching costs of the second image; and
splicing the matching costs of the first image and the matching costs of the second image to obtain the direct cost volume of the input images.
3 . The method of claim 1 , wherein processing the input images based on the second network model to obtain the associated cost volume of the input images comprises:
acquiring underlying feature matching costs of the first image and underlying feature matching costs of the second image separately based on the residual network;
grouping the underlying feature matching costs of the first image and the underlying feature matching costs of the second image separately to obtain at least one associated cost information group, wherein each of the at least one associated cost information group comprises at least one underlying feature matching cost;
updating the at least one underlying feature matching cost in the each of the at least one associated cost information group;
splicing updated underlying feature matching costs of the first image and updated underlying feature matching costs of the second image to obtain an associated cost volume; and
repeating the above grouping, updating, and splicing operations for a preset number of times, and using the obtained associated cost volume as the associated cost volume of the input images.
4 . The method of claim 3 , wherein updating the at least one underlying feature matching cost in the each of the at least one associated cost information group comprises:
calculating an average of the at least one underlying feature matching cost in the each of the at least one associated cost information group; and
replacing each of the at least one underlying feature matching cost in the each of the at least one associated cost information group with the average of the at least one underlying feature matching cost in the each of the at least one associated cost information group.
5 . The method of claim 3 , wherein a number of associated cost information groups is
group
=
{
(
epoch
+
1
)
×
4
,
group
<
C
num
4
-
1
C
num
,
group
≥
C
num
4
-
1
,
wherein group denotes the number of associated cost information groups, C num denotes a number of underlying feature matching costs, epoch denotes a current number of iterations, and a maximum value of epoch is equal to the preset number of times.
6 . The method of claim 1 , wherein determining the estimated cost of the input images according to the associated cost volume and the direct cost volume comprises:
using an average of the associated cost volume and the direct cost volume as a cost volume of the input images; and
performing cost aggregation on the cost volume based on a three-dimensional convolutional neural network to obtain the estimated cost of the input images.
7 . The method of claim 1 , wherein a disparity corresponding to the first image and the second image is
d
ˆ
=
∑
d
=
0
D
max
d
×
σ
(
-
c
d
)
,
wherein {circumflex over (d)} denotes the estimated disparity, D max denotes a maximum disparity, d denotes the disparity, σ(⋅) denotes a softmax function, and c d denotes the estimated cost.
8 . An image processing device, comprising a memory, a processor, and a computer program stored on the memory and executable by the processor, wherein the processor, when executing the computer program, performs the disparity estimation method of claim 1 .
9 . A non-transitory computer-readable storage medium, which is configured to store a computer program which, when executed by a processor, causes the processor to perform the disparity estimation method of claim 1 .