Machine-trained network for misalignment-insensitive depth perception
Some embodiments of the invention provide a novel method for training a multi-layer node network to reliably determine depth based on a plurality of input sources (e.g., cameras, microphones, etc.) that may be arranged with deviations from an ideal alignment or placement. Some embodiments train the multi-layer network using a set of inputs generated with random misalignments incorporated into the training set. In some embodiments, the training set includes (i) a synthetically generated training set based on a three-dimensional ground truth model as it would be sensed by a sensor array from different positions and with different deviations from ideal alignment and placement, and/or (ii) a training set generated by a set of actual sensor arrays augmented with an additional sensor (e.g., additional camera or time of flight measurement device such as lidar) to collect ground truth data.
1. A method for configuring a machine-trained (MT) network to account for potential deviations in locations of a plurality of image capture devices of an image producing device, each image capture device having a desired location in the image producing device, the MT network comprising a plurality of configurable parameters for training, the method comprising:
propagating a set of inputs through the MT network to generate a set of outputs, each particular input having a corresponding expected output and comprising a plurality of images of a scene each of which is captured by one of the plurality of image capture devices at an actual location in the image producing device, wherein the actual location of each particular image capture device in a set of one or more image capture devices deviates from the desired location of the particular image capture device;
computing a value for a loss function based on differences between the expected and generated outputs; and
based on the computed loss function value, back propagating a gradient of the loss function through the MT network in order to train the configurable parameters of the MT network to account for the deviations of the set of image capture devices.
2. The method of claim 1 , wherein each image in the plurality of images for each particular input is captured simultaneously.
3. The method of claim 1 , wherein:
each scene is a virtual scene;
each image capture device in the plurality of image capture devices is a virtual image capture device; and
the expected output is generated based on the virtual scene and the location of the plurality of virtual image capture devices.
4. The method of claim 1 , wherein, for each input comprising a plurality of images, the output is a disparity value for each of a plurality of pixels in at least a first image in the plurality of images, the disparity value being a measure of the number of pixels that a particular pixel shifts from one image to another.
5. The method of claim 4 , wherein the disparity values are used to calculate a depth value for each output disparity value.
6. The method of claim 1 , wherein, for each input comprising a plurality of images, the output is a depth value for each of a plurality of pixels in at least a first image in the plurality of images, each depth value indicating a distance from a point associated with the plurality of image capture devices to an object associated with a pixel of at least the first image.
7. The method of claim 1 , wherein, for each input comprising a plurality of images the output is a set of values that indicates, for a set of image capture devices in the plurality of image capture devices, a divergence from an ideal image capture device location for each image capture device in the set of image capture devices.
8. The method of claim 7 , wherein the output further comprises a set of values that indicates, for the set of image capture devices in the plurality of image capture devices, a divergence from an ideal image capture device orientation for each image capture device in the set of image capture devices.
9. The method of claim 1 , wherein the machine trained network is a convolutional neural network.
10. The method of claim 1 , wherein the plurality of image capture devices comprises image capture devices using a plurality of different lens types.
11. A non-transitory machine readable medium storing a program which when executed by at least one processing unit configures a machine-trained (MT) network to account for potential deviations in locations of a plurality of image capture devices of an image producing device, each image capture device having a desired location in the image producing device, the MT network comprising a plurality of configurable parameters for training, the program comprising sets of instructions for:
propagating a set of inputs through the MT network to generate a set of outputs, each particular input having a corresponding expected output and comprising a plurality of images of a scene each of which is captured by one of the plurality of image capture devices at an actual location in the image producing device, wherein the actual location of each particular image capture device in a set of one or more image capture devices deviates from the desired location of the particular image capture device;
computing a value for a loss function based on differences between the expected and generated outputs; and
based on the computed loss function value, back propagating a gradient of the loss function through the MT network in order to train the configurable parameters of the MT network to account for the deviations of the set of image capture devices.
12. The non-transitory machine readable medium of claim 11 , wherein each image in the plurality of images for each particular input is captured simultaneously.
13. The non-transitory machine readable medium of claim 11 , wherein
each scene is a virtual scene,
each image capture device in the plurality of image capture devices is a virtual image capture device, and
the expected output is generated based on the virtual scene and the orientation of the plurality of virtual image capture devices.
14. The non-transitory machine readable medium of claim 11 , wherein, for each input comprising a plurality of images, the output is a disparity value for each of a plurality of pixels in at least a first image in the plurality of images, the disparity value being a measure of the number of pixels that a particular pixel shifts from one image to another.
15. The non-transitory machine readable medium of claim 14 , wherein the disparity values are used to calculate a depth value for each output disparity value.
16. The non-transitory machine readable medium of claim 11 , wherein, for each input comprising a plurality of images, the output is a depth value for each of a plurality of pixels in at least a first image in the plurality of images, each depth value indicating a distance from a point associated with the plurality of image capture devices to an object associated with a pixel of at least the first image.
17. The non-transitory machine readable medium of claim 11 , wherein, for each input comprising a plurality of images the output is a set of values that indicates, for a set of image capture devices in the plurality of image capture devices, a divergence from an ideal image capture device orientation for each image capture device in the set of image capture devices.
18. The non-transitory machine readable medium of claim 16 , wherein the output further comprises a set of values that indicates, for the set of image capture devices in the plurality of image capture devices, a divergence from an ideal image capture device location for each image capture device in the set of image capture devices.
19. The non-transitory machine readable medium of claim 11 , wherein the machine trained network is a convolutional neural network.
20. The non-transitory machine readable medium of claim 11 , wherein the plurality of image capture devices comprises image capture devices using a plurality of different lens types, the plurality of different lens types comprising a rectilinear lens and a fisheye lens.