Change-aware person identification
A method for training a model, the method including: defining a primary model for identifying a class of input data based on a first characteristic of the input data; defining a secondary model for detecting a change to a second characteristic between multiple input data captured at different times; defining a forward link from an output of an intermediate layer of the secondary model to an input of an intermediate layer of the primary model; and training the primary model and the secondary model in parallel based on a training set of input data.
1. A method for training a model, the method comprising:
defining a primary model for identifying a class of input data based on a first characteristic of the input data;
defining a secondary model for detecting a change to a second characteristic between multiple input data captured at different times;
defining a forward link from an output of an intermediate layer of the secondary model to an input of an intermediate layer of the primary model;
training the primary model and the secondary model in parallel based on a training set of input data; and
after the training, discarding layers occurring after the intermediate layer of the secondary model, to produce a trained model comprising the primary model and remaining layers of the secondary model.
2. The method of claim 1 , wherein the primary model includes at least one layer for extracting a feature representation of the first characteristic and the secondary model includes at least one layer for extracting a feature representation of the second characteristic.
3. The method of claim 1 , wherein the input data is image data.
4. The method of claim 1 , wherein the first characteristic is a face and the second characteristic is clothes.
5. The method of claim 4 , wherein the forward link is from a first layer of the secondary model to the second layer of the primary model.
6. The method of claim 5 , wherein the primary model includes four layers and the secondary model includes three layers.
7. The method of claim 5 , wherein the primary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, a third BatchNow layer, and fourth Softmax layer.
8. The method of claim 7 , wherein the secondary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, and a third Softmax layer.
9. The method of claim 4 , wherein first characteristic is an output from a face convolutional neural network (CNN).
10. The method of claim 9 , wherein second characteristic is an output from a clothes convolutional neural network (CNN).
11. A non-transitory machine-readable storage medium encoded with instructions for training a model, comprising:
instructions for defining a primary model for identifying a class of input data based on a first characteristic of the input data;
instructions for defining a secondary model for detecting a change to a second characteristic between multiple input data captured at different times;
instructions for defining a forward link from an output of an intermediate layer of the secondary model to an input of an intermediate layer of the primary model; and
instructions for training the primary model and the secondary model in parallel based on a training set of input data; and
instructions for discarding layers, after the training, occurring after the intermediate layer of the secondary model, to produce a trained model comprising the primary model and remaining layers of the secondary model.
12. The non-transitory machine-readable storage medium of claim 11 , wherein the primary model includes at least one layer for extracting a feature representation of the first characteristic and the secondary model includes at least one layer for extracting a feature representation of the second characteristic.
13. The non-transitory machine-readable storage medium of claim 11 , wherein the input data is image data.
14. The non-transitory machine-readable storage medium of claim 11 , wherein the first characteristic is a face and the second characteristic is clothes.
15. The non-transitory machine-readable storage medium of claim 14 , wherein the forward link is from a first layer of the secondary model to the second layer of the primary model.
16. The non-transitory machine-readable storage medium of claim 15 , wherein the primary model includes four layers and the secondary model includes three layers.
17. The non-transitory machine-readable storage medium of claim 15 , wherein the primary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, a third BatchNow layer, and fourth Softmax layer.
18. The non-transitory machine-readable storage medium of claim 17 , wherein the secondary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, and a third Softmax layer.
19. The non-transitory machine-readable storage medium of claim 14 , wherein first characteristic is an output from a face convolutional neural network (CNN).
20. The non-transitory machine-readable storage medium of claim 19 , wherein second characteristic is an output from a clothes convolutional neural network (CNN).
21. A system for person identification from image data, the system comprising:
a processor;
a non-transitory computer readable medium that stores instructions, which when executed by the processor, causes the processor to:
define a primary model for identifying a class of input data based on a first characteristic of the input data;
define a secondary model for detecting a change to a second characteristic between multiple input data captured at different times;
define a forward link from an output of an intermediate layer of the secondary model to an input of an intermediate layer of the primary model;
train the primary model and the secondary model in parallel based on a training set of input data; and
after the primary model is trained, discard layers occurring after the intermediate layer of the secondary model, to produce a trained model comprising the primary model and remaining layers of the secondary model.
22. The system of claim 21 , wherein the primary model comprises at least one layer for extracting a feature representation of the first characteristic and the secondary model includes at least one layer for extracting a feature representation of the second characteristic.
23. The system of claim 21 , wherein the input data is image data.
24. The system of claim 21 , wherein the first characteristic is a face and the second characteristic is clothes.
25. The system of claim 24 , wherein the forward link is from a first layer of the secondary model to the second layer of the primary model.
26. The system of claim 25 , wherein the primary model includes four layers and the secondary model includes three layers.
27. The system of claim 25 , wherein the primary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, a third BatchNow layer, and fourth Softmax layer.
28. The system of claim 27 , wherein the secondary model includes a first BatchNorm+ReLU layer, a second BatchNow layer, and a third Softmax layer.
29. The system of claim 24 , wherein first characteristic is an output from a face convolutional neural network (CNN).
30. The system of claim 29 , wherein second characteristic is an output from a clothes convolutional neural network (CNN).