Three-dimensional scene reconstruction from set of two dimensional images for consumption in virtual reality
View Patent ↗To enable better sharing and preservation of immersive experiences, a graphics system reconstructs a three-dimensional scene from a set of images of the scene taken from different vantage points. The system processes each image to extract depth information therefrom and then stitches the images (both color and depth information) into a multi-layered panorama that includes at least front and back surface layers. The front and back surface layers are then merged to remove redundancies and create connections between neighboring pixels that are likely to represent the same object, while removing connections between neighboring pixels that are not. The resulting layered panorama with depth information can be rendered using a virtual reality (VR) system, a mobile device, or other computing and display platforms using standard rendering techniques, to enable three-dimensional viewing of the scene.
1. A method comprising:
receiving a plurality of input images of a scene taken from different vantage points;
processing the plurality of input images to generate a sparse reconstruction representation of the scene, the sparse reconstruction representation including a sparse point cloud specifying locations of a plurality of points that correspond to three-dimensional locations of surfaces of objects in the scene;
based in part on the sparse reconstruction representation, processing the plurality of input images to generate respective dense reconstruction representations of each of the plurality of input images, the respective dense reconstruction representations each including respective depth images for the plurality of input images, the depth images including color and depth information;
projecting front surfaces of the depth images using a forward depth test to generate a plurality of front-warped images;
projecting back surfaces of the depth images using an inverted depth test to generate a plurality of back-warped images;
stitching the front-warped images and the back-warped images to generate a two-layer panorama having a front surface panorama and a back surface panorama; and
fusing the front surface panorama and the back surface panorama in the two-layer panorama to generate the three-dimensional image comprising a multi-layered geometric mesh suitable for rendering the scene in the three-dimensional space.
2. The method of claim 1 , wherein processing the plurality of input images to generate the sparse reconstruction representation comprises:
applying a surface-from-motion algorithm to the plurality of input images.
3. The method of claim 1 , wherein processing the plurality of input images to generate the respective dense reconstruction representations comprises:
generating a near envelope prior that assigns a cost to estimated depth values in front of a near envelope; and
applying a multi-view stereo processing algorithm to estimate the depth values based on a cost function including the near envelope prior.
4. The method of claim 3 , wherein generating the near envelope prior comprises:
identifying anchor pixels in the plurality of input images that have high confidence depth estimates;
propagating the depth estimates of the anchor pixels to other pixels in the plurality of input images to generate approximate depth maps;
filtering the approximate depth maps to determine a near envelope; and
generating the near envelope prior based on the depth estimates and the near envelope.
5. The method of claim 1 , wherein stitching the front-warped images and the back-warped images to generate a two-layer panorama comprises:
stitching a depth panorama using depth values from the front-warped images;
stitching the front surface panorama using color values from the front-warped images and stitched depth values from the depth panorama;
stitching the back surface panorama using color values from the back-warped images and the stitched depth values from the depth panorama; and
combining the front surface panorama and the back surface panorama into the two-layer panorama.
6. The method of claim 1 , wherein fusing the front surface panorama and the back surface panorama comprises:
removing background pixels from the back surface panorama that match corresponding foreground pixels in the front surface panorama;
storing connections between neighboring pixels meeting a threshold similarity in depth and color information; and
hallucinating color and depth information in missing pixel locations.
7. The method of claim 1 , further comprising:
generating a normal map for the multi-layered geometric mesh, the normal map estimating for each pixel, an angle normal to a surface depicted by the pixel.
8. The method of claim 7 , wherein generating the normal map comprises:
generating a base normal map from depth values in the three-dimensional image;
generating a detailed normal from luminance values in the three-dimensional image; and
transforming the detailed normal map onto the base normal map to generate a combined normal map.
9. The method of claim 1 , wherein the plurality of input images having varying levels of overlap and orientation changes.
10. A non-transitory computer-readable storage medium storing instructions, the instructions when executed by a processor causing the processor to perform steps including:
receiving a plurality of input images of a scene taken from different vantage points;
processing the plurality of input images to generate a sparse reconstruction representation of the scene, the sparse reconstruction representation including a sparse point cloud specifying locations of a plurality of points that correspond to three-dimensional locations of surfaces of objects in the scene;
based in part on the sparse reconstruction representation, processing the plurality of input images to generate respective dense reconstruction representations of each of the plurality of input images, the respective dense reconstruction representations each including respective depth images for the plurality of input images, the depth images including color and depth information;
projecting front surfaces of the depth images using a forward depth test to generate a plurality of front-warped images;
projecting back surfaces of the depth images using an inverted depth test to generate a plurality of back-warped images;
stitching the front-warped images and the back-warped images to generate a two-layer panorama having a front surface panorama and a back surface panorama;
fusing the front surface panorama and the back surface panorama in the two-layer panorama to generate the three-dimensional image comprising a multi-layered geometric mesh suitable for rendering the scene in the three-dimensional space.
11. The non-transitory computer-readable storage medium of claim 10 , wherein processing the plurality of input images to generate the sparse reconstruction representation comprises:
applying a surface-from-motion algorithm to the plurality of input images.
12. The non-transitory computer-readable storage medium of claim 10 , wherein processing the plurality of input images to generate the respective dense reconstruction representations comprises:
generating a near envelope prior that assigns a cost to estimated depth values in front of a near envelope; and
applying a multi-view stereo processing algorithm to estimate the depth values based on a cost function including the near envelope prior.
13. The non-transitory computer-readable storage medium of claim 12 , wherein generating the near envelope prior comprises:
identifying anchor pixels in the plurality of input images that have high confidence depth estimates;
propagating the depth estimates of the anchor pixels to other pixels in the plurality of input images to generate approximate depth maps;
filtering the approximate depth maps to determine a near envelope; and
generating the near envelope prior based on the depth estimates and the near envelope.
14. The non-transitory computer-readable storage medium of claim 10 , wherein stitching the front-warped images and the back-warped images to generate a two-layer panorama comprises:
stitching a depth panorama using depth values from the front-warped images;
stitching the front surface panorama using color values from the front-warped images and stitched depth values from the depth panorama;
stitching the back surface panorama using color values from the back-warped images and the stitched depth values from the depth panorama; and
combining the front surface panorama and the back surface panorama into the two-layer panorama.
15. The non-transitory computer-readable storage medium of claim 10 , wherein fusing the front surface panorama and the back surface panorama comprises:
removing background pixels from the back surface panorama that match corresponding foreground pixels in the front surface panorama;
storing connections between neighboring pixels meeting a threshold similarity in depth and color information; and
hallucinating color and depth information in missing pixel locations.
16. The non-transitory computer-readable storage medium of claim 10 , wherein the instructions when executed by processor further cause the processor to perform steps including:
generating a normal map for the multi-layered geometric mesh, the normal map estimating for each pixel, an angle normal to a surface depicted by the pixel.
17. The non-transitory computer-readable storage medium of claim 16 , wherein generating the normal map comprises:
generating a base normal map from depth values in the three-dimensional image;
generating a detailed normal from luminance values in the three-dimensional image; and
transforming the detailed normal map onto the base normal map to generate a combined normal map.
18. The non-transitory computer-readable storage medium of claim 10 , wherein the plurality of input images having varying levels of overlap and orientation changes.
19. A system comprising:
a processor; and
a non-transitory computer-readable storage medium storing instruction for generating a three-dimensional image, the instructions when executed by a processor causing the processor to perform steps including:
receiving a plurality of input images of a scene taken from different vantage points;
processing the plurality of input images to generate a sparse reconstruction representation of the scene, the sparse reconstruction representation including a sparse point cloud specifying locations of a plurality of points that correspond to three-dimensional locations of surfaces of objects in the scene;
based in part on the sparse reconstruction representation, processing the plurality of input images to generate respective dense reconstruction representations of each of the plurality of input images, the respective dense reconstruction representations each including respective depth images for the plurality of input images, the depth images including color and depth information;
projecting front surfaces of the depth images using a forward depth test to generate a plurality of front-warped images;
projecting back surfaces of the depth images using an inverted depth test to generate a plurality of back-warped images;
stitching the front-warped images and the back-warped images to generate a two-layer panorama having a front surface panorama and a back surface panorama;
fusing the front surface panorama and the back surface panorama in the two-layer panorama to generate the three-dimensional image comprising a multi-layered geometric mesh suitable for rendering the scene in the three-dimensional space.
20. The system of claim 19 , wherein stitching the front-warped images and the back-warped images to generate a two-layer panorama comprises:
stitching a depth panorama using depth values from the front-warped images;
stitching the front surface panorama using color values from the front-warped images and stitched depth values from the depth panorama;
stitching the back surface panorama using color values from the back-warped images and the stitched depth values from the depth panorama; and
combining the front surface panorama and the back surface panorama into the two-layer panorama.