Modifying an image based on identifying a feature

1. An apparatus comprising:

a camera that captures an image;

a processor;

a memory that stores code executable by the processor to:

detect a feature in the image using a convolutional neural network trained on a feature training set, wherein the feature is selected from the group consisting of one or more fingers, one or more fingertips, one or more gloved fingers, one or more gloved fingertips, a hand, a gloved hand, an instrument, and a tool;

place the feature at a center of a displayed image;

determine an intent from a voice command using a recurrent neural network trained on motions of the feature for the image, wherein intent is selected from the group comprising a specified zoom, a maximum zoom, a zoom in, a zoom out, following the feature, a pan left, a pan right, a pan up, and a pan down, and the recurrent neural network is a Long Short Term Memory (LSTM) neural network that receives a plurality of temporal instances of the feature; and

modify the displayed image based on the intent, wherein the plurality of temporal instances of the feature are each scaled to one of two or more window sizes and each window size is a one dimensional kernel of a specified length.

2. A method comprising:

detecting, by use of a processor, a feature in an image using a convolutional neural network trained on a feature training set, wherein the feature is selected from the group consisting of one or more fingers, one or more fingertips, one or more gloved fingers, one or more gloved fingertips, a hand, a gloved hand, an instrument, and a tool;

placing the feature at a center of a displayed image;

determining an intent from a voice command using a recurrent neural network trained on motions of the feature for the image, wherein intent is selected from the group comprising a specified zoom, a maximum zoom, a zoom in, a zoom out, following the feature, a pan left, a pan right, a pan up, and a pan down, and the recurrent neural network is a Long Short Term Memory (LSTM) neural network that receives a plurality of temporal instances of the feature; and

modifying the displayed image based on the intent, wherein the plurality of temporal instances of the feature are each scaled to one of two or more window sizes and each window size is a one dimensional kernel of a specified length.

3. A program product comprising a computer readable storage medium that stores code executable by a processor, the executable code comprising code to:

detect a feature in an image using a convolutional neural network trained on a feature training set, wherein the feature is selected from the group consisting of one or more fingers, one or more fingertips, one or more gloved fingers, one or more gloved fingertips, a hand, a gloved hand, an instrument, and a tool;

place the feature at a center of a displayed image;

determine an intent from a voice command using a recurrent neural network trained on motions of the feature for the image, wherein intent is selected from the group comprising a specified zoom, a maximum zoom, a zoom in, a zoom out, following the feature, a pan left, a pan right, a pan up, and a pan down, and the recurrent neural network is a Long Short Term Memory (LSTM) neural network that receives a plurality of temporal instances of the feature; and

modify the displayed image based on the intent, wherein the plurality of temporal instances of the feature are each scaled to one of two or more window sizes and each window size is a one dimensional kernel of a specified length.

4. The apparatus of claim 1 , wherein the convolutional neural network generates a temporal slice that is flattened across a time interval and the flattened temporal slice is input into the recurrent neural network.

5. The apparatus of claim 1 , wherein the camera captures a wide field-of-view and the displayed image is modified by selecting a portion of the field-of-view to be presented.

6. The method of claim 2 , wherein the convolutional neural network generates a temporal slice that is flattened across a time interval and the flattened temporal slice is input into the recurrent neural network.

7. The method of claim 2 , wherein a camera captures a wide field-of-view and the displayed image is modified by selecting a portion of the field-of-view to be presented.

8. The program product of claim 3 , wherein the convolutional neural network generates a temporal slice that is flattened across a time interval and the flattened temporal slice is input into the recurrent neural network.

9. The program product of claim 3 , wherein a camera captures a wide field-of-view and the displayed image is modified by selecting a portion of the field-of-view to be presented.