Eye tracking in near-eye displays
Techniques for tracking eye movement in an augmented reality system identify a plurality of base images of an object or a portion thereof. A search image may be generated based at least in part upon at least some of the plurality of base images. A deep learning result may be generated at least by performing a deep learning process on a base image using a neural network in a deep learning mode. A captured image may be localized at least by performing an image registration process on the captured image and the search image using a Kalman filter model and the deep learning result.
1. A method, comprising:
determining a search image that is constructed based at least in part upon a plurality of template image frames;
capturing, using a spatial computing headset, a plurality of captured image frames of an object or a portion thereof;
performing an image registration process that registers two or more captured image frames of the plurality of captured image frames in the search image using a deep network;
tracking, using the spatial computing headset, the movement of the object based at least in part upon respective results of registering the plurality of captured image frames in the search image; and
training the deep network using at least some captured image frames of the plurality of captured image frames, training the deep network comprising
identifying a captured image frame of the at least some captured image frames captured by the spatial computing headset,
extracting, using the deep network, a capture image frame feature from a first region in the captured image frame and a search image feature from a second region in the search image,
converting the captured image frame feature into a plurality of features that comprises a pair of a first feature and a second feature for the deep network,
providing the first feature converted from the captured image frame feature to a classification subnetwork of the deep network, and
providing the second feature converted from the captured image frame feature to a regression subnetwork of the deep network.
2. The method of claim 1 , wherein tracking the movement of the object is performed by the spatial computing headset, without capturing glint reflected from the object in response to an input light pattern.
3. The method of claim 1 , wherein training the deep network is accomplished without using regularization techniques.
4. The method of claim 1 , training the deep network further comprising:
producing, by the classification subnetwork, a first output data structure at least by convolving, at the classification subnetwork, the first feature converted from the captured image frame feature and search image feature from the search image; and
producing, by the regression subnetwork, a second output data structure at least by convolving, at the regression subnetwork, the second feature converted from the captured image frame feature and search image feature from the search image.
5. The method of claim 3 , training the deep network further comprising:
determining, by the classification subnetwork, whether the first region belongs to a target region or a non-target region based at least in part upon the first output data structure; and
predicting, by the regression subnetwork, a position refinement for the region that is determined to belong to the target region or the non-target region based at least in part upon the second output data structure.
6. The method of claim 1 , wherein the spatial computing headset uses a position of the object or a feature or a portion of the object, instead of geometric attributes of bounding boxes for the object or the feature or the portion of the object in tracking the movement of the object.
7. A system, comprising:
a spatial computing headset comprising:
a processor,
a scanning fiber assembly, and
memory storing thereupon a sequence of instructions, which, when executed by the processor, causes the processor to perform a set of acts, the set of acts comprising:
determining a search image that is constructed based at least in part upon a plurality of template image frames;
capturing, using a spatial computing headset, a plurality of captured image frames of an object or a portion thereof;
performing an image registration process that registers two or more captured image frames of the plurality of captured image frames in the search image using a deep network;
tracking, using the spatial computing headset, the movement of the object based at least in part upon respective results of registering the plurality of captured image frames in the search image; and
training the deep network using at least some captured image frames of the plurality of captured image frames, training the deep network comprising
identifying a captured image frame of the at least some captured image frames captured by the spatial computing headset,
extracting, using the deep network, a capture image frame feature from a first region in the captured image frame and a search image feature from a second region in the search image,
converting the captured image frame feature into a plurality of features that comprises a pair of a first feature and a second feature for the deep network,
providing the first feature converted from the captured image frame feature to a classification subnetwork of the deep network, and
providing the second feature converted from the captured image frame feature to a regression subnetwork of the deep network.
8. The system of claim 7 , wherein tracking the movement of the object is performed by the spatial computing headset, without capturing glint reflected from the object in response to an input light pattern.
9. The system of claim 7 , wherein the set of acts comprises training the deep network, training the deep network comprising:
producing, by the classification subnetwork, a first output data structure at least by convolving, at the classification subnetwork, the first feature converted from the captured image frame feature and search image feature from the search image; and
producing, by the regression subnetwork, a second output data structure at least by convolving, at the regression subnetwork, the second feature converted from the captured image frame feature and search image feature from the search image.
10. The system of claim 9 , wherein the set of acts comprises training the deep network, training the deep network comprising:
determining, by the classification subnetwork, whether the first region belongs to a target region or a non-target region based at least in part upon the first output data structure; and
predicting, by the regression subnetwork, a position refinement for the region that is determined to belong to the target region or the non-target region based at least in part upon the second output data structure.
11. A non-transitory computer-readable medium storing thereupon instructions which, when executed by a microprocessor, causes the microprocessor to perform a set of acts, the set of acts comprising:
determining a search image that is constructed based at least in part upon a plurality of template image frames;
capturing, using a spatial computing headset, a plurality of captured image frames of an object or a portion thereof;
performing an image registration process that registers two or more captured image frames of the plurality of captured image frames in the search image using a deep network;
tracking, using the spatial computing headset, the movement of the object based at least in part upon respective results of registering the plurality of captured image frames in the search image; and
training the deep network using at least some captured image frames of the plurality of captured image frames, training the deep network comprising
identifying a captured image frame of the at least some captured image frames captured by the spatial computing headset,
extracting, using the deep network, a capture image frame feature from a first region in the captured image frame and a search image feature from a second region in the search image,
converting the captured image frame feature into a plurality of features that comprises a pair of a first feature and a second feature for the deep network,
providing the first feature converted from the captured image frame feature to a classification subnetwork of the deep network, and
providing the second feature converted from the captured image frame feature to a regression subnetwork of the deep network,
wherein tracking the movement of the object is performed by the spatial computing headset, without capturing glint reflected from the object in response to an input light pattern.
12. The non-transitory computer-readable medium of claim 11 , wherein the set of acts comprises training the deep network, training the deep network comprising:
producing, by the classification subnetwork, a first output data structure at least by convolving, at the classification subnetwork, the first feature converted from the captured image frame feature and search image feature from the search image; and
producing, by the regression subnetwork, a second output data structure at least by convolving, at the regression subnetwork, the second feature converted from the captured image frame feature and search image feature from the search image.
13. The system of claim 12 , wherein the set of acts comprises training the deep network, training the deep network comprising:
determining, by the classification subnetwork, whether the first region belongs to a target region or a non-target region based at least in part upon the first output data structure; and
predicting, by the regression subnetwork, a position refinement for the region that is determined to belong to the target region or the non-target region based at least in part upon the second output data structure.