Network architecture for monocular depth estimation and object detection

1. A depth system, comprising:

one or more processors; and

a memory communicably coupled to the one or more processors and storing:

a network module including instructions that, when executed by the one or more processors, cause the one or more processors to:

generate, using a backbone of a combined network, a feature map at multiple scales from an input image;

decode, using a top-down pathway of the combined network, the feature map to provide features at the multiple scales;

generate, using a head of the combined network, a depth map from the features for a scene depicted in the input image and bounding boxes identifying objects in the input image, wherein the head includes multiple sub-heads for each of the multiple scales that perform 3D object detection, 2D object detection, depth estimation, and classification;

train, in a first stage, the combined network by using a supervised depth loss derived from the depth map; and

train, in a second stage, the combined network by using the bounding boxes and ground-truth data to compute a detection loss.

2. The depth system of claim 1 , wherein the network module includes instructions to decode including instructions to provide, using lateral connections between the backbone and the top-down pathway, the multiple scales of the feature map in addition to an output of a prior level from within the top-down pathway.

3. The depth system of claim 1 , wherein the network module includes instructions to generate the feature map including instructions to generate the feature map at the multiple scales as a feature hierarchy, and

wherein the network module includes instructions to generate the feature map to encode features of the input image to provide a common reference for generating the depth map and the bounding boxes.

4. The depth system of claim 1 , wherein the network module includes instructions to generate the depth map and the bounding boxes including instructions to use the head among separate layers of the top-down pathway at the multiple scales to generate the bounding boxes at the multiple scales and the depth map at one of the multiple scales.

5. The depth system of claim 1 , wherein the input image is a monocular image in RGB.

6. A non-transitory computer-readable medium including instructions that, when executed by one or more processors, cause the one or more processors to:

generate, using a backbone of a combined network, a feature map at multiple scales from an input image;

decode, using a top-down pathway of the combined network, the feature map to provide features at the multiple scales; and

generate, using a head of the combined network, a depth map from the features for a scene depicted in the input image and bounding boxes identifying objects in the input image, wherein the head includes multiple sub-heads for each of the multiple scales that perform 3D object detection, 2D object detection, depth estimation, and classification;

train, in a first stage, the combined network by using a supervised depth loss derived from the depth map; and

train, in a second stage, the combined network by using the bounding boxes and ground-truth data to compute a detection loss.

7. The non-transitory computer-readable medium of claim 6 , wherein the instructions to decode include instructions to provide, using lateral connections between the backbone and the top-down pathway, the multiple scales of the feature map in addition to an output of a prior level from within the top-down pathway.

8. The non-transitory computer-readable medium of claim 6 , wherein the instructions to generate the feature map including instructions to generate the feature map at the multiple scales as a feature hierarchy, and

wherein the instructions to generate the feature map encode features of the input image to provide a common reference for generating the depth map and the bounding boxes.

9. A method, comprising:

generating, using a backbone of a combined network, a feature map at multiple scales from an input image;

decoding, using a top-down pathway of the combined network, the feature map to provide features at the multiple scales; and

generating, using a head of the combined network, a depth map from the features for a scene depicted in the input image and bounding boxes identifying objects in the input image, wherein the head includes multiple sub-heads for each of the multiple scales that perform 3D object detection, 2D object detection, depth estimation, and classification;

training, in a first stage, the combined network by using a supervised depth loss derived from the depth map; and

training, in a second stage, the combined network by using the bounding boxes and ground-truth data to compute a detection loss.

10. The method of claim 9 , wherein decoding includes providing, using lateral connections between the backbone and the top-down pathway, the multiple scales of the feature map in addition to an output of a prior level from within the top-down pathway.

11. The method of claim 9 , wherein generating the feature map includes generating the feature map at the multiple scales as a feature hierarchy, and

wherein generating the feature map encodes features of the input image to provide a common reference for generating the depth map and the bounding boxes.

12. The method of claim 9 , wherein generating the depth map and the bounding boxes includes using the head among separate layers of the top-down pathway at the multiple scales to generate the bounding boxes at the multiple scales and the depth map at one of the multiple scales.

13. The method of claim 9 , further comprising:

providing the depth map and the bounding boxes to cause navigation of a device according to the depth map and the bounding boxes.