Method and system for processing an image by determining rotation hypotheses
A system and a method for processing an image include inputting the image to a neural network configured to: obtain a plurality of feature maps, each feature map having a respective resolution and a respective depth, perform a classification on each feature map to deliver, for each feature map: the type of at least one object visible on the image, the position and shape in the image of at least one two-dimensional bounding box surrounding the at least one object, a plurality of rotation hypotheses for the at least one object.
1. A method for processing an image using a processor, the method comprising:
inputting, by a processor, an image to a neural network;
obtaining, by the processor, a plurality of feature maps, each feature map having a respective resolution and a respective depth;
performing, by the processor, a classification on each feature map to deliver, wherein for each feature map:
the type of at least one object visible on the image,
the position and shape in the image of at least one two-dimensional bounding box surrounding the at least one object,
a plurality of rotation hypotheses for the at least one object, and
a distance, between the at least one object and an image acquisition module having acquired the image, determined using the diagonal length of the at least one bounding box.
2. The method according to claim 1 , further comprising performing a classification of the plurality of rotation hypotheses to obtain an estimated rotation.
3. The method of claim 2 , wherein the classification of the plurality of rotation hypotheses comprises a rejection of outliers or a clustering.
4. The method according to claim 2 , wherein the estimated rotation is expressed as a quaternion.
5. The method according to claim 2 , further comprising performing a principal component analysis on the rotation hypotheses so as to determine whether the estimated rotation exhibits an ambiguity.
6. The method of claim 1 , further comprising performing a classification of the plurality of rotation hypotheses to obtain an estimated rotation and using the estimated rotation and the distance, between the at least one object and the image acquisition module, to estimate the 6D pose of the at least one object.
7. The method according to claim 1 , comprising a preliminary training of the neural network, the preliminary training comprising inputting a template image in which at least one three-dimensional template object has been rendered to the neural network.
8. The method of claim 7 , wherein the preliminary training further comprises:
obtaining, using the neural network:
the type of at least one template object visible on the template image,
the position and shape in the image of at least one two-dimensional bounding box surrounding the at least one template object,
a plurality of rotation hypotheses for the at least one template object, and minimizing L in the following equation:
L
(
Pos
,
Neg
)
=
∑
b
∈
Neg
L
class
+
∑
b
∈
Pos
(
L
class
+
α
L
fit
+
β
M
(
f
θ
(
I
)
,
p
)
)
in which:
M
(
f
θ
(
I
)
,
p
)
=
(
1
-
ϵ
M
M
-
1
)
M
^
(
f
θ
(
I
)
,
p
)
+
ϵ
M
-
1
∑
j
=
1
M
L
(
f
θ
(
j
)
(
I
)
,
p
)
M
ˆ
(
f
θ
(
I
)
,
p
)
=
min
j
=
1
,
…
,
M
L
(
f
θ
(
j
)
(
I
)
,
p
)
and:
I is the template image,
Pos represents the bounding boxes which have an area which encloses a predefined fraction of area of the template object,
Neg represents the bounding boxes which have an area which does not enclose a predefined fraction of area of the template object,
b is a bounding box belonging to either Pos or Neg,
L class is the loss associated with errors in the attribution of class,
L fit is the loss associated with errors the location of the corners of bounding boxes,
L view is the loss associated with errors in the attribution of viewpoints,
L inplane is the loss associated with errors in the attribution of in-plane rotations,
α, β and ϵ are predetermined coefficients,
M is the number of rotation hypotheses,
I is the template image
p is the true rotation of the at least one object,
ƒ θ (j) (I) is the function which delivers the rotation hypothesis j,
ƒ θ (I) is the function which delivers the rotation hypotheses ƒ θ (j) (I),
L(ƒ θ (j) (I),p) is a loss function of ƒ θ (I) for a known rotation p.
9. The method of claim 8 , wherein during the training, the neural network is updated using a rotation hypothesis θ* determined as:
θ
*
=
arg
min
θ
∑
i
=
1
N
M
(
f
θ
(
I
i
)
,
p
i
)
wherein p i are N rotations associated with images I t which have a similar image for the at least one template object.
10. A non-transitory computer readable medium storing a computer program including instructions for executing the steps of the method according to claim 1 when the program is executed by a computer.
11. A recording medium readable by a computer and having recorded thereon a computer program including instructions for executing the steps of a method according to claim 1 .
12. A system for processing an image, the system comprising:
a processor executing a neural network to which an image can be inputted, the processor being configured to:
obtain a plurality of feature maps, each feature map having a respective resolution and a respective depth, and
perform a classification on each feature map to deliver, for each feature map:
the type of at least one object visible on the image,
the position and shape in the image of at least one two-dimensional bounding box surrounding the at least one object,
a plurality of rotation hypotheses for the at least one object, and
a distance, between the object and an image acquisition module having acquired the image, determined using the diagonal length of the at least one bounding box.