Video annotation using deep network architectures

1. A method comprising:

receiving, by a processing device of a content sharing platform, a video content;

selecting at least one video frame from the video content, wherein the at least one video frame covers a spatial area at a first resolution;

subsampling the at least one video frame to generate a first representation of the at least one video frame, wherein the first representation is at a second resolution that is lower than the first resolution;

selecting, at the first resolution, a sub-region of the at least one video frame to generate a second representation of the at least one video frame, wherein the sub-region covers a spatial area that is smaller than the spatial area covered by the at least one video frame; and

executing a convolutional neuron network, using the first representation as a first input of the convolutional neuron network and using the second representation as a second input of the convolutional neuron network, to generate an annotation for the video content.

2. The method of claim 1 , wherein the second representation is a fovea representation that is at a same spatial sampling rate as the at least one video frame.

3. The method of claim 1 , wherein the at least one video frame is a single frame.

4. The method of claim 1 , wherein the at least one video frame includes one of two consecutive video frames or at least two non-consecutive video frames.

5. The method of claim 1 , wherein the convolutional neuron network includes at least one convolution layer, at least one pooling layer, and a connected neuron network.

6. The method of claim 5 , wherein a first convolution layer and a first pooling layer are applied to a first number of video frames, and a second convolution layer and a second pooling layer are applied to a second number of video frames, and wherein the first number is different from the second number.

7. The method of claim 5 , wherein an earlier layer of the convolutional neuron network is applied to a higher number of video frames than a later layer of the convolutional neuron network.

8. The method of claim 1 , further comprising making the video content searchable according to the annotation.

9. A non-transitory machine-readable storage medium storing instructions which, when executed, cause a processing device to perform operations comprising:

receiving a video content;

selecting at least one video frame from the video content, wherein the at least one video frame covers a spatial area at a first resolution;

subsampling the at least one video frame to generate a first representation of the at least one video frame, wherein the first representation is at a second resolution that is lower than the first resolution;

selecting, at the first resolution, a sub-region of the at least one video frame to generate a second representation of the at least one video frame, wherein the sub-region covers a spatial area that is smaller than the spatial area covered by the at least one video frame; and

executing a convolutional neuron network, using the first representation as a first input of the convolutional neuron network and using second representation as a second input of the convolutional neuron network, to generate an annotation for the video content.

10. The machine-readable storage medium of claim 9 , wherein the second representation is a fovea representation that is at a same spatial sampling rate as the at least one video frame.

11. The machine-readable storage medium of claim 9 , wherein the at least one video frame is a single frame.

12. The machined-readable storage medium of claim 9 , wherein the at least one video frame includes one of at least two consecutive video frames or at least two non-consecutive video frames.

13. The machine-readable storage medium of claim 9 , wherein the convolutional neuron network includes at least one convolution layer, at least one pooling layer, and a connected neuron network.

14. The machine-readable storage medium of claim 11 , wherein a first convolution layer and a first pooling layer are applied to a first number of video frames, and a second convolution layer and a second pooling layer are applied to a second number of video frames, and wherein the first number is different from the second number.

15. A system comprising:

a memory; and

a processor, operatively coupled to the memory, to:

receive a video content;

select at least one video frame from the video content, wherein the at least one video frame covers a spatial area at a first resolution;

subsample the at least one video frame to generate a first representation of the at least one video frame, wherein the first representation is at a second resolution that is lower than the first resolution;

select, at the first resolution, a sub-region of the at least one video frame to generate a second representation of the at least one video frame, wherein the sub-region covers a spatial area that is smaller than the spatial area covered by the at least one video frame; and

execute a convolutional neuron, using the first representation as a first input of the convolutional neuron network and using the second representation as a second input of the convolutional neuron network, to generate an annotation for the video content.

16. The system of claim 15 , wherein the second representation is a fovea representation that is at a same spatial sampling rate as the at least one video frame.

17. The system of claim 15 , wherein the convolutional neuron network includes at least one convolution layer, at least one pooling layer, and a connected neuron network.

18. The user device of claim 17 , wherein a first convolution layer and a first pooling layer are applied to a first number of video frames, and a second convolution layer and a second pooling layer are applied to a second number of video frames, and wherein the first number is different from the second number.