Cross-domain few-shot video classification with optical-flow semantics
Methods and systems for video processing include extracting flow features and appearance features from frames of a video stream. The flow features are processed using a flow model that is trained on a first set of training data. An output of the flow model is processed using a sub-network that is trained on the first set of training data and a second set of domain-specific training data to generate a flow parameter. The appearance features are processed using an appearance model that is trained on the first set of training data and that further processes the appearance features using the flow parameter, to classify the frames of the video stream. An action is performed responsive to the classified frames.
1 . A computer-implemented method for video processing, comprising:
extracting flow features and appearance features from frames of a video stream;
processing the flow features using a flow model that is trained on a first set of training data;
processing an output of the flow model using a sub-network that is trained on the first set of training data and a second set of domain-specific training data to generate a flow parameter;
processing the appearance features using an appearance model that is trained on the first set of training data and that further processes the appearance features using the flow parameter, to classify the frames of the video stream; and
performing an action responsive to the classified frames.
2 . The method of claim 1 , wherein processing the appearance features includes a convolutional network that processes the appearance features in parallel with the flow parameter.
3 . The method of claim 2 , wherein processing the appearance features includes adding a product of flow parameter and the appearance features to an output of the convolutional network.
4 . The method of claim 1 , wherein the appearance model includes a series of blocks, each of which includes a convolutional network and a respective instance of the flow parameter.
5 . The method of claim 1 , wherein the flow features characterize motion information relating to the frames of the video stream.
6 . The method of claim 1 , wherein the appearance features characterize pixel value information from the frames of the video stream.
7 . The method of claim 1 , wherein performing the action includes an action selected from the group consisting of generating additional information relating to the classified frames and performing a security action relating to the classified frames.
8 . A system for video processing, comprising:
a hardware processor; and
a memory that stores a computer program which, when executed by the hardware processor, causes the hardware processor to:
extract flow features and appearance features from frames of a video stream;
process the flow features using a flow model that is trained on a first set of training data;
process an output of the flow model using a sub-network that is trained on the first set of training data and a second set of domain-specific training data to generate a flow parameter;
process the appearance features using an appearance model that is trained on the first set of training data and that further processes the appearance features using the flow parameter, to classify the frames of the video stream; and
perform an action responsive to the classified frames.
9 . The system of claim 8 , wherein the computer program further causes the hardware processor to use a convolutional network that processes the appearance features in parallel with the flow parameter.
10 . The system of claim 9 , wherein the computer program further causes the hardware processor to add a product of flow parameter and the appearance features to an output of the convolutional network.
11 . The system of claim 8 , wherein the appearance model includes a series of blocks, each of which includes a convolutional network and a respective instance of the flow parameter.
12 . The system of claim 8 , wherein the flow features characterize motion information relating to the frames of the video stream.
13 . The system of claim 8 , wherein the appearance features characterize pixel value information from the frames of the video stream.
14 . The system of claim 8 , wherein the computer program further causes the hardware processor to perform action selected from the group consisting of generating additional information relating to the classified frames and performing a security action relating to the classified frames.
15 . A computer-implemented method for training a neural network model, comprising:
training a flow model to process flow features of frames of a video stream, a sub-network to process outputs of the flow model, and an appearance model to process appearance features of the frames with a flow parameter from the sub-network, using a shared first set of training data; and
tuning the sub-network while the flow model and the appearance model are held fixed, using a second set of domain-specific training data.
16 . The method of claim 15 , wherein the appearance model includes a convolutional network that processes the appearance features in parallel with the flow parameter.
17 . The method of claim 16 , wherein the appearance model adds a product of flow parameter and the appearance features to an output of the convolutional network.
18 . The method of claim 15 , wherein the appearance model includes a series of blocks, each of which includes a convolutional network and a respective instance of the flow parameter.
19 . The method of claim 15 , further comprising extracting flow features from the frames to characterize motion information relating to the frames.
20 . The method of claim 15 , further comprising extracting appearance features from the frames to characterize pixel value information from the frames.