Densifying sparse depth maps
A system and techniques that use one or more machine learning models to predict a dense depth map (e.g., of depth values for all pixels or at least more pixels than a sparse estimation source (e.g., SLAM)). In some implementations, the machine learning model includes two sub models (e.g., neural networks). The first machine learning model predicts computer vision data such as semantic labels and surface normal directions from an input image. This computer vision data will be used to add to or otherwise improve sparse depth data. Specifically, a second machine learning model takes the semantic labels and surface normal directions from and sparse depth data (e.g., 3D points) from a sparse point estimation source (e.g., SLAM) as inputs and outputs a depth map. The output depth map effectively densities the initial depth data (e.g., from SLAM) by providing depth data for additional pixels of the image.
1. A method, comprising:
at an electronic device having a processor:
obtaining an image of a physical setting from an image capture device;
producing a semantic prediction and a surface normal prediction using a first machine learning model with input comprising the image;
producing depth data using a depth estimator on the image; and
outputting a depth map using a second machine learning model with input comprising the semantic prediction, the surface normal prediction, and the depth data.
2. The method of claim 1 , wherein the depth map is denser than the depth data, the depth map providing depth estimates for more portions of the image than the depth data.
3. The method of claim 1 , wherein the depth map is denser than the depth data, the depth map providing depth estimates for more pixels of the image than the depth data.
4. The method of claim 1 , wherein the depth map provides a depth estimate for each pixel of the image.
5. The method of claim 1 , wherein the depth estimator is a simultaneous localization and mapping (SLAM) technique.
6. The method of claim 5 , wherein the depth data is determined by projecting three dimensional (3D) points determined by the SLAM technique based on:
a pose of the image capture device during capture of the image; and
focal length or distortion parameters of the image capture device.
7. The method of claim 1 , wherein the first machine learning model comprises multiple sub-models, the sub-models comprising a semantic prediction model and a surface normal prediction model.
8. The method of claim 1 , wherein outputting the depth map comprises displaying the depth map.
9. The method of claim 1 , wherein the second model is trained by:
obtaining sample images from the image capture device;
producing training depth maps using a structure-from-motion (SFM) technique on the sample images;
producing semantic predictions and surface normal predictions using the first machine learning model with input comprising the sample images;
producing depth data sets using the depth estimator on the sample images;
outputting depth maps using the second machine learning model with input comprising the semantic predictions, the surface normal predictions, and the depth data sets; and
adjusting the second machine learning model based on comparing the training depth maps and the output depth maps.
10. The method of claim 9 , wherein the comparing excludes pixels that are not predicted by the SFM technique.
11. A system comprising:
a non-transitory computer-readable storage medium; and
a processor coupled to the non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium comprises program instructions that, when executed on the processor, cause the system to perform operations comprising:
obtaining an image of a physical setting from an image capture device;
producing a semantic prediction and a surface normal prediction using a first machine learning model with input comprising the image;
producing depth data using a depth estimator on the image; and
outputting a depth map using a second machine learning model with input comprising the semantic prediction, the surface normal prediction, and the depth data.
12. The system of claim 11 , wherein the depth map provides depth estimates for more pixels of the image than the depth data.
13. The system of claim 11 , wherein the depth map provides a depth estimate for each pixel of the image.
14. The system of claim 11 , wherein the depth estimator is a simultaneous localization and mapping (SLAM) technique.
15. The system of claim 14 , wherein the depth data is determined by projecting three dimensional (3D) points determined by the SLAM technique based on:
a pose of the image capture device during capture of the image; and
focal length or distortion parameters of the image capture device.
16. The system of claim 11 , wherein the second model is trained by:
obtaining sample images from the image capture device;
producing training depth maps using a structure-from-motion (SFM) technique on the sample images;
producing semantic predictions and surface normal predictions using the first machine learning model with input comprising the sample images;
producing depth data sets using the depth estimator on the sample images;
outputting depth maps using the second machine learning model with input comprising the semantic predictions, the surface normal predictions, and the depth data sets; and
adjusting the second machine learning model based on comparing the training depth maps and the output depth maps, wherein the comparing excludes pixels that are not predicted by the SFM technique.
17. A non-transitory computer-readable storage medium, storing program instructions computer-executable on a computer to perform operations comprising:
obtaining an image of a physical setting from an image capture device;
producing a semantic prediction and a surface normal prediction using a first machine learning model with input comprising the image;
producing depth data using a depth estimator on the image; and
outputting a depth map using a second machine learning model with input comprising the semantic prediction, the surface normal prediction, and the depth data.
18. The non-transitory computer-readable storage medium of claim 17 , wherein the depth map provides depth estimates for more pixels of the image than the depth data.
19. The non-transitory computer-readable storage medium of claim 17 , wherein the depth data is determined by projecting three dimensional (3D) points determined by a SLAM technique based on:
a pose of the image capture device during capture of the image; and
focal length or distortion parameters of the image capture device.
20. The non-transitory computer-readable storage medium of claim 17 , wherein the second model is trained by:
obtaining sample images from the image capture device;
producing training depth maps using a structure-from-motion (SFM) technique on the sample images;
producing semantic predictions and surface normal predictions using the first machine learning model with input comprising the sample images;
producing depth data sets using the depth estimator on the sample images;
outputting depth maps using the second machine learning model with input comprising the semantic predictions, the surface normal predictions, and the depth data sets; and
adjusting the second machine learning model based on comparing the training depth maps and the output depth maps, wherein the comparing excludes pixels that are not predicted by the SFM technique.