Depth determination system and method

1 . A method comprising:

receiving an image sampled by an image sensor with a known focal length;

predicting a standard disparity map from the image using a standard disparity model; and

determining a depth map corresponding to the image using the known focal length and the standard disparity map;

wherein the standard disparity model comprises a trained neural network, wherein training the standard disparity model comprises:

determining a plurality of training images, each associated with a training focal length and an independently-sampled training depth map; and

for each training image:

determining a target standard disparity map based on the respective training focal length and the respective training depth map;

predicting a predicted standard disparity map based on the training image using the standard disparity model; and

updating the standard disparity model based on a comparison between the predicted standard disparity map and the target standard disparity map.

2 . The method of claim 1 , wherein the standard disparity map comprises a standard disparity at each of a set of pixels, wherein determining the depth map comprises determining a depth for each of the set of pixels by scaling an inverse standard disparity at each pixel of the standard disparity map with the known focal length.

3 . The method of claim 1 , wherein the standard disparity map comprises a standard disparity for more than a majority of pixels in the image.

4 . The method of claim 1 , wherein the known focal length is in pixels.

5 . The method of claim 1 , wherein the plurality of training images is sampled using a heterogeneous set of focal lengths.

6 . The method of claim 1 , wherein the training image is associated with an independently-sampled training surface normal map, wherein training the standard disparity model further comprises:

extracting a set of target vanishing lines from the training surface normal map;

independently predicting a set of predicted vanishing lines based on the training image using the standard disparity model; and

updating the standard disparity model based on a comparison between the set of target vanishing lines and the set of predicted vanishing lines.

7 . The method of claim 1 , wherein training the standard disparity model further comprises:

independently predicting a set of predicted vanishing lines based on the training image using the standard disparity model; and

updating the standard disparity model based on a comparison between a gradient of the predicted standard disparity map and a surface normal proxy determined from the set of predicted vanishing lines.

8 . The method of claim 1 , wherein the image consists of a monocular image.

9 . The method of claim 1 , wherein the standard disparity model is used to predict a second standard disparity map for a second image associated with a second focal length different from the known focal length; wherein a second depth map corresponding to the second image is determined from the second standard disparity map and the second focal length.

10 . The method of claim 1 , wherein the image sensor is mounted to a vehicle, the method further comprising controlling the vehicle using the depth map.

11 . The method of claim 1 , wherein the standard disparity map is further predicted based on a set of sparse depth values associated with the image.

12 . A system, comprising:

a processing system;

a non-transitory computer readable medium storing software instructions that, when executed by a processing system, cause the processing system to perform a method comprising:

receiving an image captured by an image sensor with a first focal length;

predicting a normalized inverse depth map for the image using a trained normalized inverse depth model; and

determining a dense depth map for the image based on the first focal length and the normalized inverse depth map;

wherein the normalized inverse depth model comprises a trained neural network, wherein training the normalized inverse depth model comprises:

determining a set of training images, wherein different training images in the set of training images are associated with different focal lengths and different depth maps;

for each training image:

determining a target normalized inverse depth map based on the respective focal length and the respective depth map;

predicting a predicted normalized inverse depth map from the training image using the normalized inverse depth model; and

updating parameters of the normalized inverse depth model based on a comparison between the predicted normalized inverse depth map and the target normalized inverse depth map.

13 . The system of claim 12 , wherein the first focal length is in pixels.

14 . The system of claim 12 , further comprising a vehicle comprising the image sensor, wherein the vehicle is controlled based on the normalized inverse depth map.

15 . The system of claim 14 , wherein the vehicle further comprises a second image sensor with a second focal length different from the first focal length, wherein a second normalized inverse depth map is predicted from a second image sampled by the second image sensor using the normalized inverse depth model, and a second depth map is determined from the second normalized inverse depth map using the second focal length.

16 . The system of claim 12 , wherein the image consists essentially of a monocular image.

17 . The system of claim 12 , wherein training the normalized inverse depth model further comprises, for each training image:

predicting a surface normal proxy map from the training image using the normalized inverse depth model;

determining a normalized inverse depth gradient map from the predicted normalized inverse depth map; and

updating the normalized inverse depth model by enforcing a consistency constraint between the normalized inverse depth gradient map and the surface normal proxy map.

18 . The system of claim 12 , wherein the normalized inverse depth map is further predicted using a set of sparse depth proxy priors.