Disparity-based depth refinement using confidence information and stereoscopic depth information
Systems and techniques are provided for generating disparity information from two or more images. For example, a process can include obtaining first disparity information corresponding to a pair of images, the pair of images including a first image of a scene and a second image of the scene. The process can include obtaining confidence information associated with the first disparity information. The process can include processing, using a machine learning network, the first disparity information and the confidence information to generate second disparity information corresponding to the pair of images. The process can include combining, based on the confidence information, the first disparity information with the second disparity information to generate a refined disparity map corresponding to the pair of images.
1 . An apparatus for generating disparity information from two or more images, the apparatus comprising:
at least one memory; and
at least one processor coupled to the at least one memory, the at least one processor being configured to:
obtain first disparity information corresponding to a pair of images, the pair of images including a first image of a scene and a second image of the scene, wherein the first disparity information comprises an estimated disparity map between the first and second images;
obtain confidence information associated with the first disparity information;
generate second disparity information corresponding to the pair of images, wherein the second disparity information comprises an initial refinement of the first disparity information based on the confidence information, and wherein the second disparity information is generated as an output prediction of a disparity refinement machine learning network configured to process the first disparity information and the confidence information as inputs; and
combine, based on the confidence information, the first disparity information with the second disparity information to generate a refined disparity map corresponding to the pair of images, wherein the refined disparity map comprises a weighted combination of the estimated disparity map and the initial refinement according to the confidence information.
2 . The apparatus of claim 1 , wherein the pair of images is a stereoscopic image pair including a left view of the scene and a right view of the scene.
3 . The apparatus of claim 1 , wherein, to generate the refined disparity map, the at least one processor is configured to:
fuse the first disparity information with the confidence information to generate a first fused disparity map;
fuse the second disparity information with a complement of the confidence information to generate a second fused disparity map; and
combine the first fused disparity map and the second fused disparity map to generate the refined disparity map.
4 . The apparatus of claim 3 , wherein:
the confidence information includes a confidence map comprising a plurality of locations having a resolution, each location of the plurality of locations including a confidence value representing a respective confidence value of a disparity prediction having a corresponding location in the first disparity information; and
the complement of the confidence information comprises a plurality of locations having the resolution, each location of the plurality of locations including a value determined as one minus the respective confidence value having a corresponding location in the confidence map.
5 . The apparatus of claim 1 , wherein:
the first disparity information is the estimated disparity map and wherein the estimated disparity map includes a plurality of estimated disparity values, each respective estimated disparity value of the plurality of estimated disparity values associated with a particular location of a plurality of locations in the estimated disparity map; and
the second disparity information is a predicted disparity map including a plurality of predicted disparity values, each respective predicted disparity value of the plurality of predicted disparity values associated with a particular location of a plurality of locations in the predicted disparity map.
6 . The apparatus of claim 5 , wherein:
the plurality of locations in the estimated disparity map is the same as the plurality of locations in the predicted disparity map; and
the respective predicted disparity value associated with the particular location in the predicted disparity map corresponds to the respective estimated disparity value associated with the particular location in the estimated disparity map.
7 . The apparatus of claim 5 , wherein the at least one processor is configured to:
generate the respective predicted disparity value associated with the particular location based on using the machine learning network to refine the respective estimated disparity value associated with the particular location.
8 . The apparatus of claim 1 , wherein:
the first disparity information comprises a depth-from-stereo (DFS) disparity map generated based on feature matching between the first image and the second image; and
the second disparity information comprises a predicted disparity map generated based on the DFS disparity map, the confidence information, and at least one of the first image or the second image.
9 . The apparatus of claim 1 , wherein the machine learning network is a neural network trained to generate a refined disparity map corresponding to an input disparity map.
10 . The apparatus of claim 1 , wherein:
each of the first image and the second image includes a plurality of pixels having a resolution;
the first disparity information includes the estimated disparity map for the pair of images, the estimated disparity map comprising a plurality of locations having the resolution;
the confidence information includes a confidence map corresponding to the estimated disparity map, the confidence map comprising a plurality of locations having the resolution; and
the second disparity information includes a predicted disparity map for the pair of images, the predicted disparity map comprising a plurality of locations having the resolution.
11 . The apparatus of claim 10 , wherein:
each location of a plurality of locations in the refined disparity map includes a fused disparity value representing a weighted combination of a respective estimated disparity value having a corresponding location in the estimated disparity map and a respective predicted disparity value having a corresponding location in the predicted disparity map; and
the weighted combination is based on a respective confidence value having a corresponding location in the confidence map.
12 . The apparatus of claim 11 , wherein the at least one processor is configured to determine the weighted combination based on weighting the respective estimated disparity value by the respective confidence value and weighting the respective predicted disparity value by one minus the respective confidence value.
13 . The apparatus of claim 1 , wherein, to obtain the confidence information, the at least one processor is configured to:
obtain a plurality of features corresponding to the pair of images;
process the plurality of features and the first disparity information to determine one or more locations included in the first disparity information corresponding to a respective one or more occluded regions of the pair of images; and
generate a respective confidence value for each location of a plurality of locations included in the confidence information, wherein the respective confidence value is based on whether each location corresponds to the one or more occluded regions.
14 . The apparatus of claim 13 , wherein, to obtain the confidence information, the at least one processor is further configured to:
determine a left-to-right disparity map between the first and second images of the pair of images, using the first image as an anchor image for the first disparity map;
determine a right-to-left disparity map between the first and second images of the pair of images, using the second image as an anchor image for the second disparity map; and
process the left-to-right disparity map and the right-to-left disparity map to generate left-right consistency information indicative of one or more valid locations included in the first disparity information and one or more invalid locations included in the first disparity information.
15 . The apparatus of claim 14 , wherein the at least one processor is further configured to:
generate the respective confidence value further based on whether each location corresponds to a valid or invalid location included in the first disparity information.
16 . A method generating disparity information from two or more images, comprising:
obtaining first disparity information corresponding to a pair of images, the pair of images including a first image of a scene and a second image of the scene, wherein the first disparity information comprises an estimated disparity map between the first and second images;
obtaining confidence information associated with the first disparity information;
generating second disparity information corresponding to the pair of images, wherein the second disparity information comprises an initial refinement of the first disparity information based on the confidence information, and wherein the second disparity information is generated as an output prediction of a disparity refinement machine learning network configured to process the first disparity information and the confidence information as inputs; and
combining, based on the confidence information, the first disparity information with the second disparity information to generate a refined disparity map corresponding to the pair of images, wherein the refined disparity map comprises a weighted combination of the estimated disparity map and the initial refinement according to the confidence information.
17 . The method of claim 16 , wherein the pair of images is a stereoscopic image pair including a left view of the scene and a right view of the scene.
18 . The method of claim 16 , wherein generating the refined disparity map comprises:
fusing the first disparity information with the confidence information to generate a first fused disparity map;
fusing the second disparity information with a complement of the confidence information to generate a second fused disparity map; and
combining the first fused disparity map and the second fused disparity map to generate the refined disparity map.
19 . The method of claim 18 , wherein:
the confidence information includes a confidence map comprising a plurality of locations having a resolution, each location of the plurality of locations including a confidence value representing a respective confidence value of a disparity prediction having a corresponding location in the first disparity information; and
the complement of the confidence information comprises a plurality of locations having the resolution, each location of the plurality of locations including a value determined as one minus the respective confidence value having a corresponding location in the confidence map.
20 . The method of claim 16 , wherein:
the first disparity information is the estimated disparity map and wherein the estimated disparity map includes a plurality of estimated disparity values, each respective estimated disparity value of the plurality of estimated disparity values associated with a particular location of a plurality of locations in the estimated disparity map; and
the second disparity information is a predicted disparity map including a plurality of predicted disparity values, each respective predicted disparity value of the plurality of predicted disparity values associated with a particular location of a plurality of locations in the predicted disparity map.
21 . The method of claim 20 , wherein:
the plurality of locations in the estimated disparity map is the same as the plurality of locations in the predicted disparity map; and
the respective predicted disparity value associated with the particular location in the predicted disparity map corresponds to the respective estimated disparity value associated with the particular location in the estimated disparity map.
22 . The method of claim 20 , further comprising:
generating the respective predicted disparity value associated with the particular location based on using the machine learning network to refine the respective estimated disparity value associated with the particular location.
23 . The method of claim 16 , wherein:
the first disparity information comprises a depth-from-stereo (DFS) disparity map generated based on feature matching between the first image and the second image; and
the second disparity information comprises a predicted disparity map generated based on the DFS disparity map, the confidence information, and at least one of the first image or the second image.
24 . The method of claim 16 , wherein the machine learning network is a neural network trained to generate a refined disparity map corresponding to an input disparity map.
25 . The method of claim 16 , wherein:
each of the first image and the second image includes a plurality of pixels having a resolution;
the first disparity information includes the estimated disparity map for the pair of images, the estimated disparity map comprising a plurality of locations having the resolution;
the confidence information includes a confidence map corresponding to the estimated disparity map, the confidence map comprising a plurality of locations having the resolution; and
the second disparity information includes a predicted disparity map for the pair of images, the predicted disparity map comprising a plurality of locations having the resolution.
26 . The method of claim 25 , wherein:
each location of a plurality of locations in the refined disparity map includes a fused disparity value representing a weighted combination of a respective estimated disparity value having a corresponding location in the estimated disparity map and a respective predicted disparity value having a corresponding location in the predicted disparity map; and
the weighted combination is based on a respective confidence value having a corresponding location in the confidence map.
27 . The method of claim 26 , further comprising determining the weighted combination based on weighting the respective estimated disparity value by the respective confidence value and weighting the respective predicted disparity value by one minus the respective confidence value.
28 . The method of claim 16 , wherein obtaining the confidence information comprises:
obtaining a plurality of features corresponding to the pair of images;
processing the plurality of features and the first disparity information to determine one or more locations included in the first disparity information corresponding to a respective one or more occluded regions of the pair of images; and
generating a respective confidence value for each location of a plurality of locations included in the confidence information, wherein the respective confidence value is based on whether each location corresponds to the one or more occluded regions.
29 . The method of claim 28 , wherein obtaining the confidence information comprises:
determining a left-to-right disparity map between the first and second images of the pair of images, using the first image as an anchor image for the first disparity map;
determining a right-to-left disparity map between the first and second images of the pair of images, using the second image as an anchor image for the second disparity map; and
processing the left-to-right disparity map and the right-to-left disparity map to generate left-right consistency information indicative of one or more valid locations included in the first disparity information and one or more invalid locations included in the first disparity information.
30 . The method of claim 29 , further comprising:
generating the respective confidence value further based on whether each location corresponds to a valid or invalid location included in the first disparity information.