RGB-D scene labeling with multimodal recurrent neural networks
Described herein are systems and methods for multimodal recurrent network processing. In an embodiment, a system for evaluating multimodal data comprising a multimodal data input and a multimodal processing module is described. The multimodal data input may comprise the multimodal data, the multimodal data may comprise a first modality and a second modality. The multimodal processing module may be configured to receive the multimodal data comprising the first modality and the second modality; evaluate the first modality using a first recursive neural network comprising a first transformation matrix; evaluate the second modality using a second recursive neural network comprising the first transformation matrix; and determine an output based, at least in part, on evaluating the first modality and the second modality.
1. A system for evaluating multimodal data comprising:
a multimodal data input comprising the multimodal data, the multimodal data comprising a first modality and a second modality; and
a multimodal processing module configured to:
receive the multimodal data comprising the first modality and the second modality;
evaluate the first modality using a first recursive neural network comprising a first transformation matrix;
evaluate the second modality using a second recursive neural network comprising the first transformation matrix, wherein the first transformation matrix shares memory of the first modality and the second modality between the first recursive network and the second recursive network to inform the first recursive network and the second recursive network of inter-correlations between the first modality and the second modality; and
determine an output based, at least in part, on evaluating the first modality and the second modality.
2. The system of claim 1 , wherein the first recursive neural network comprises a first hidden layer and the second recursive neural network comprises a second hidden layer, and wherein the multimodal processing module is further configured to:
evaluate the first hidden layer using the first transformation matrix; and
evaluate the second hidden layer using the first transformation matrix, wherein the first transformation matrix shares learned weights between the first recursive neural network and the second recursive neural network.
3. The system of claim 2 , wherein the first recursive neural network comprises a first output layer and the second recursive neural network comprises a second output layer, and wherein the multimodal processing module is further configured to determine the first output layer and the second output layer using a second transformation matrix that shares learned weights of the first output layer and the second output layer to provide memory of cross-correlations between the first modality and the second modality.
4. The system of claim 3 , wherein the multimodal processing module is further configured to determine the output based, at least in part, upon a summation of the first output layer and the second output layer.
5. The system of claim 4 , wherein the multimodal processing module is further configured to evaluate the multimodal data based, at least in part, on an equation set, as follows:
h
t
m
=
∅
(
U
m
x
t
m
+
W
h
t
-
1
m
+
b
h
m
)
y
t
m
=
σ
(
V
h
t
m
+
b
y
m
)
o
t
=
∑
m
=
1
M
w
m
y
t
m
wherein W represents the first transformation matrix and V represents the second transformation matrix across the first modality and the second modality, wherein h m t represents a hidden layer and y m t represents an output layer for a respective modality, and wherein o t represents a final output via weighted summation for the first modality and the second modality.
6. The system of claim 1 , wherein the multimodal data input comprises an RGB-D image, and wherein the first modality comprises an RGB image and the second modality comprises a depth image.
7. The system of claim 1 , wherein the multimodal processing module is configured to implement hidden layers of the first recursive neural network and the second recursive neural network using directed acyclic graphs that define the hidden layers at separate vertices with the first transformation matrix and a second transformation matrix sharing memory of the first modality and the second modality between corresponding ones of the vertices of the first recursive neural network and the second recursive neural network.
8. An automobile comprising:
a RGB-D camera configured to:
capture a RGB-D image; and
output a RGB modality and a depth modality; and
an image processing module in communication with the RGB-D camera, the image processing module configured to:
receive the RGB modality and the depth modality;
evaluate the RGB modality using a first recursive neural network comprising a first transformation matrix;
evaluate the depth modality using a second recursive neural network comprising the first transformation matrix, wherein the first transformation matrix shares memory of the first modality and the second modality between the first recursive network and the second recursive network to inform the first recursive network and the second recursive network of inter-correlations between the first modality and the second modality; and
determine a prediction comprising predictions of objects in the RGB-D image.
9. The automobile of claim 8 , wherein the first recursive neural network comprises a first hidden layer and the second recursive neural network comprises a second hidden layer, and wherein the image processing module is further configured to:
evaluate the first hidden layer using the first transformation matrix; and
evaluate the second hidden layer using the first transformation matrix, wherein the first transformation matrix shares learned weights between the first recursive neural network and the second recursive neural network.
10. The automobile of claim 9 , wherein the first recursive neural network comprises a first output layer and the second recursive neural network comprises a second output layer, and wherein the image processing module is further configured to determine the first output layer and the second output layer using a second transformation matrix that shares learned weights of the first output layer and the second output layer to provide memory of cross-correlations between the RGB modality and the depth modality.
11. The automobile of claim 10 , wherein the image processing module is further configured to determine the prediction based, at least in part, upon a summation of the first output layer and the second output layer.
12. The automobile of claim 11 , wherein the image processing module is further configured to evaluate the RGB-D image based, at least in part, on an equation set, as follows:
h
t
m
=
∅
(
U
m
x
t
m
+
W
h
t
-
1
m
+
b
h
m
)
y
t
m
=
σ
(
V
h
t
m
+
b
y
m
)
o
t
=
∑
m
=
1
M
w
m
y
t
m
wherein W represents the first transformation matrix and V represents the second transformation matrix across the first modality and the second modality, wherein h m t represents a hidden layer and y m t represents an output layer for a respective modality, and wherein o t represents a final output via weighted summation for the RGB modality and the depth modality.
13. The automobile of claim 8 , wherein the image processing module is further configured to implement hidden layers of the first recursive neural network and the second recursive neural network using directed acyclic graphs that define the hidden layers at separate vertices with the first transformation matrix and a second transformation matrix sharing memory of the RGB modality and the depth modality between corresponding ones of the vertices of the first recursive neural network and the second recursive neural network.
14. A method for evaluating multimodal data, the method comprising:
receiving the multimodal data comprising a first modality and a second modality;
evaluating the first modality using a first recursive neural network comprising a first transformation matrix;
evaluating the second modality using a second recursive neural network comprising the first transformation matrix, wherein the first transformation matrix shares memory of the first modality and the second modality between the first recursive network and the second recursive network to inform the first recursive network and the second recursive network of inter-correlations between the first modality and the second modality; and
determining an output based, at least in part, on evaluating the first modality and the second modality.
15. The method of claim 14 , further comprising:
evaluating a first hidden layer of the first recursive neural network using the first transformation matrix; and
evaluating a second hidden layer of the second recursive neural network using the first transformation matrix, wherein the first transformation matrix shares learned weights between the first recursive neural network and the second recursive neural network.
16. The method of claim 15 , further comprising determining a first output layer of the first recursive neural network and a second output layer of the second recursive neural network using a second transformation matrix that shares learned weights of the first output layer and the second output layer to provide memory of cross-correlations between the first modality and the second modality.
17. The method of claim 16 , further comprising determining the output based, at least in part, upon a summation of the first output layer and the second output layer.
18. The method of claim 17 , wherein evaluating the first modality and evaluating the second modality is based, at least in part, on an equation set, as follows:
h
t
m
=
∅
(
U
m
x
t
m
+
W
h
t
-
1
m
+
b
h
m
)
y
t
m
=
σ
(
V
h
t
m
+
b
y
m
)
o
t
=
∑
m
=
1
M
w
m
y
t
m
wherein W represents the first transformation matrix and V represents the second transformation matrix across the first modality and the second modality, wherein h m t represents a hidden layer and y m t represents an output layer for a respective modality, and wherein o t represents a final output via weighted summation for the first modality and the second modality.
19. The method of claim 14 , wherein the multimodal data comprises an RGB-D image, and wherein the first modality comprises an RGB image and the second modality comprises a depth image.
20. The method of claim 14 , further comprising implementing hidden layers of the first recursive neural network and the second recursive neural network using directed acyclic graphs that define the hidden layers at separate vertices with the first transformation matrix and a second transformation matrix sharing memory of the first modality and the second modality between corresponding ones of the vertices of the first recursive neural network and the second recursive neural network.
21. The method of claim 14 , wherein the multimodal data further comprises a plurality of additional modalities, the method further comprising:
evaluating each of the plurality of additional modalities using a plurality of additional recursive neural networks each of the plurality of additional recursive neural networks comprising the first transformation matrix.