Camera agnostic depth network

1. A method for monocular depth/pose estimation in a camera agnostic network, the method comprising:

projecting lifted 3D points onto an image plane according to a predicted ray vector based on a monocular depth model, a monocular pose model, and a camera center according to the camera agnostic network; and

predicting a warped target image from a predicted depth map of the monocular depth model, a ray surface of the predicted ray vector, and a projection of the lifted 3D points according to the camera agnostic network.

2. The method of claim 1 , further comprising:

training the monocular depth model and the monocular pose model to learn monocular depth estimation and monocular pose estimation based on a target image and one or more context images from monocular video captured by the camera agnostic network; and

lifting 3D points from image pixels of the target image according to the one or more context images.

3. The method of claim 2 , in which training comprises self-supervised learning of an ego-motion model to predict rigid transformations between the target image and the one or more context images captured by the camera agnostic network, in which the one or more context images represent adjacent frames of monocular video.

4. The method of claim 2 , in which training comprises incorporating intrinsic parameters and extrinsic parameters regarding a plurality of different cameras into a camera agnostic configuration of the monocular depth model and the monocular pose model.

5. The method of claim 4 , in which the extrinsic parameters define a location and orientation of the plurality of different cameras with respect to a world frame, and the intrinsic parameters enable a mapping between camera coordinates and pixel coordinates in an image frame.

6. The method of claim 1 , further comprising estimating a pose transformation from a target image to one or more context images based on predicted rigid transformations between the target image and the one or more context images captured by the camera agnostic network, in which the one or more context images represent adjacent frames of monocular video.

7. The method of claim 1 , further comprising estimating a trajectory of an ego vehicle based on the warped target image.

8. The method of claim 7 , further comprising planning a trajectory of the ego vehicle according to an estimated trajectory of the ego vehicle.

9. The method of claim 1 , further comprising:

scaling the predicted ray vector according to a predicted depth; and

offsetting the predicted ray vector by the camera center.

10. A non-transitory computer-readable medium having program code recorded thereon for monocular depth/pose estimation in a camera agnostic network, the program code being executed by a processor and comprising:

program code to project lifted 3D points onto an image plane according to a predicted ray vector based on a monocular depth model, a monocular pose model, and a camera center according to the camera agnostic network; and

program code to predict a warped target image from a predicted depth map of the monocular depth model, a ray surface of the predicted ray vector, and a projection of the lifted 3D points according to the camera agnostic network.

11. The non-transitory computer-readable medium of claim 10 , further comprising:

program code to train the monocular depth model and the monocular pose model to learn monocular depth estimation and monocular pose estimation based on a target image and one or more context images from monocular video captured by the camera agnostic network;

program code to lift 3D points from image pixels of the target image according to the one or more context images.

12. The non-transitory computer-readable medium of claim 11 , in which training comprises self-supervised learning of an ego-motion model to predict rigid transformations between the target image and the one or more context images captured by the camera agnostic network, in which the one or more context images represent adjacent frames of monocular video.

13. The non-transitory computer-readable medium of claim 11 , in which the program code to train comprises program code to incorporate intrinsic parameters and extrinsic parameters regarding a plurality of different cameras into a camera agnostic configuration of the monocular depth model and the monocular pose model, and

in which the extrinsic parameters define a location and orientation of the plurality of different cameras with respect to a world frame, and the intrinsic parameters enable a mapping between camera coordinates and pixel coordinates in an image frame.

14. The non-transitory computer-readable medium of claim 10 , further comprising estimating a pose transformation from a target image to one or more context images based on predicted rigid transformations between the target image and the one or more context images captured by the camera agnostic network, in which the one or more context images represent adjacent frames of monocular video.

15. The non-transitory computer-readable medium of claim 10 , further comprising:

program code to estimate a trajectory of an ego vehicle based on the warped target image; and

program code to plan a trajectory of the ego vehicle according to an estimated trajectory of the ego vehicle.

16. The non-transitory computer-readable medium of claim 10 , further comprising:

program code to scale the predicted ray vector according to a predicted depth; and

program code to offset the predicted ray vector by the camera center.

17. A system for monocular depth/pose estimation in a camera agnostic network, the system comprising:

a pose network to project lifted 3D points onto an image plane according to a predicted ray vector based on a monocular depth model, a monocular pose model, and a camera center according to the camera agnostic network; and

a view synthesis block to predict a warped target image from a predicted depth map of the monocular depth model, a ray surface of the predicted ray vector, and a projection of the lifted 3D points according to the camera agnostic network.

18. The system of claim 17 , further comprising an ego perception module to estimate a trajectory of an ego vehicle based on the warped target image.

19. The system of claim 18 , further comprising a planner module to plan a trajectory of the ego vehicle according to an estimated trajectory of the ego vehicle.

20. The system of claim 17 , further comprising:

a depth network to train the monocular depth model and the monocular pose model to learn monocular depth estimation and monocular pose estimation based on a target image and one or more context images from monocular video captured by the camera agnostic network,

in which the pose network is further to lift 3D points from image pixels of the target image according to the one or more context images.