Image depth prediction with wavelet decomposition
A depth prediction model for predicting a depth map from an input image is disclosed. The depth prediction model leverages wavelet decomposition to minimize computations. The depth prediction model comprises a plurality of encoding layers, a coarse prediction layer, a plurality of decoding layers, and a plurality of inverse discrete wavelet transforms (IDWTs). The encoding layers are configured to input the image and to downsample the image into feature maps including a coarse feature map. The coarse depth prediction layer is configured to input the coarse feature map and to output a coarse depth map. The decoding layers are configured to input the feature maps and to predict wavelet coefficients based on the feature maps. The IDWTs are configured to upsample the coarse depth map based on the predicted wavelet coefficients to the final depth map at the same resolution as the input image.
1 . A method comprising:
receiving an image captured by a camera on a client device;
applying a depth prediction model to the image to generate a depth map based on the image, the depth prediction model comprising:
a plurality of encoding layers configured to input the image and to downsample the image into one or more feature maps including a coarse feature map,
a coarse depth prediction layer configured to input the coarse feature map and to output a coarse depth map based on the coarse feature map,
a plurality of decoding layers configured to input the one or more feature maps and to predict wavelet coefficients based on the one or more feature maps, wherein:
a first decoding layer of the plurality is configured to:
input the coarse feature map at a first resolution, predict wavelet coefficients at the first resolution, and output a first feature map at a second resolution that is higher than the first resolution; and
a second decoding layer of the plurality is configured to:
input the first feature map at the second resolution that is output by the first decoding layer,
predict sparse wavelet coefficients at the second resolution, and
output a second feature map at a third resolution that is higher than the second resolution; and
a plurality of inverse discrete wavelet transforms configured to upsample the
coarse depth map based on the predicted wavelet coefficients;
generating a virtual element based on the depth map; and
displaying the virtual element with the image on an electronic display of the client device.
2 . The method of claim 1 , wherein each encoding layer is configured to downsample by a common factor, and each decoding layer is configured to upsample by the common factor.
3 . The method of claim 1 , wherein a first encoding layer is configured to downsample by a first factor, and a second encoding layer is configured to downsample by a second factor that is different than the first factor.
4 . The method of claim 1 , wherein each decoding layer is configured to predict at least one of: Haar wavelet coefficients, Daubechies wavelet coefficients, and LeGall-Tabatai 5/3 wavelet coefficients.
5 . The method of claim 1 , wherein the second decoding layer is further configured to:
concatenate the first feature map at the second resolution with a third feature map at the
second resolution that is output by one of the encoding layers, and input the first feature map concatenated with the third feature map.
6 . The method of claim 1 , wherein the second decoding layer is configured to predict the sparse wavelet coefficients by applying a binary mask that is generated based on the wavelet coefficients at the first resolution.
7 . The method of claim 1 , wherein a number of encoding layers is equal to a number of decoding layers.
8 . The method of claim 1 , wherein the depth map has the same resolution as the image.
9 . The method of claim 1 , wherein the depth prediction model is a machine-learned model trained using a plurality of training images with ground truth depth maps.
10 . A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform operations comprising:
receiving an image captured by a camera on a client device;
applying a depth prediction model to the image to generate a depth map based on the image, the depth prediction model comprising:
a plurality of encoding layers configured to input the image and to downsample the image into one or more feature maps including a coarse feature map,
a coarse depth prediction layer configured to input the coarse feature map and to output a coarse depth map based on the coarse feature map,
a plurality of decoding layers configured to input the one or more feature maps and to predict wavelet coefficients based on the one or more feature maps, wherein:
a first decoding layer of the plurality is configured to:
input the coarse feature map at a first resolution, predict wavelet coefficients at the first resolution, and output a first feature map at a second resolution that is higher than the first resolution; and
a second decoding layer of the plurality is configured to:
input the first feature map at the second resolution that is output by the first decoding layer,
predict sparse wavelet coefficients at the second resolution, and
output a second feature map at a third resolution that is higher than the second resolution; and
a plurality of inverse discrete wavelet transforms configured to upsample the coarse depth map based on the predicted wavelet coefficients;
generating a virtual element based on the depth map; and
displaying the virtual element with the image on an electronic display of the client device.
11 . The non-transitory computer-readable storage medium of claim 10 , wherein each encoding layer is configured to downsample by a common factor, and each decoding layer is configured to upsample by the common factor.
12 . The non-transitory computer-readable storage medium of claim 10 , wherein a first encoding layer is configured to downsample by a first factor, and a second encoding layer is configured to downsample by a second factor that is different than the first factor.
13 . The non-transitory computer-readable storage medium of claim 10 , wherein each decoding layer is configured to predict at least one of: Haar wavelet coefficients, Daubechies wavelet coefficients, and LeGall-Tabatai 5/3 wavelet coefficients.
14 . The non-transitory computer-readable storage medium of claim 10 , wherein the second decoding layer is further configured to:
concatenate the first feature map at the second resolution with a third feature map at the second resolution that is output by one of the encoding layers, and
input the first feature map concatenated with the third feature map.
15 . The non-transitory computer-readable storage medium of claim 10 , wherein the second decoding layer is configured to predict the sparse wavelet coefficients by applying a binary mask that is generated based on the wavelet coefficients at the first resolution.
16 . The non-transitory computer-readable storage medium of claim 10 , wherein a number of encoding layers is equal to a number of decoding layers.
17 . The non-transitory computer-readable storage medium of claim 10 , wherein the depth map has the same resolution as the image.
18 . The non-transitory computer-readable storage medium of claim 10 , wherein the depth prediction model is a machine-learned model trained using a plurality of training images with ground truth depth maps.
19 . A method comprising:
receiving an image captured by a camera on an autonomous agent;
applying a depth prediction model to the image to generate a depth map based on the image, the depth prediction model comprising:
a plurality of encoding layers configured to input the image and to downsample the image into one or more feature maps including a coarse feature map,
a coarse depth prediction layer configured to input the coarse feature map and to output a coarse depth map based on the coarse feature map,
a plurality of decoding layers configured to input the one or more feature maps and to predict wavelet coefficients based on the one or more feature maps, wherein:
a first decoding layer of the plurality is configured to:
input the coarse feature map at a first resolution, predict wavelet coefficients at the first resolution, and output a first feature map at a second resolution that is higher than the first resolution; and
a second decoding layer of the plurality is configured to:
input the first feature map at the second resolution that is output by the first decoding layer,
predict sparse wavelet coefficients at the second resolution, and
output a second feature map at a third resolution that is higher than the second resolution; and
a plurality of inverse discrete wavelet transforms configured to upsample the coarse depth map based on the predicted wavelet coefficients;
generating navigation instructions based on the depth map; and
navigating the autonomous agent based on the navigation instructions.