Methods and systems for using compact object image data to construct a machine learning model for pose estimation of an object

1. A method comprising:

accessing, by a model construction system, object image data representative of one or more images depicting an object having a plurality of labeled keypoint features;

generating, by the model construction system and based on the object image data, a training target dataset including a plurality of training target images, wherein a particular training target image of the plurality of training target images is generated by:

selecting a background image from a set of background images distinct from the object image data,

manipulating a depiction of the object represented within the object image data, and

overlaying the manipulated depiction of the object onto the selected background image together with an indication of labeled keypoint features for the manipulated depiction; and

training, by the model construction system based on the training target dataset, a machine learning model to recognize and estimate a pose of the object when the object is depicted in input images analyzed using the trained machine learning model.

2. The method of claim 1 , wherein the manipulating of the depiction of the object includes performing at least one of:

a rotation operation to rotate the depiction of the object with respect to two or three spatial dimensions; or

a scaling operation to change a size of the depiction of the object.

3. The method of claim 1 , wherein the manipulating of the depiction of the object includes performing a cropping operation to simulate an occlusion of a portion of the depiction of the object.

4. The method of claim 1 , wherein:

the set of background images corresponds to a background image library configured for use in generating training target datasets for a variety of objects including the object and other objects; and

the background image library comprises background images that are captured independently from image capture associated with the object.

5. The method of claim 1 , wherein:

the particular training target image is further generated by applying, subsequent to the overlaying of the manipulated depiction of the object onto the selected background image, an image processing operation to the particular training target image; and

the image processing operation changes at least one of a contrast attribute, a color attribute, or a saturation attribute of the particular training target image.

6. The method of claim 1 , further comprising:

recognizing, based on the machine learning model, the object as depicted within a particular input image captured by an augmented reality presentation device; and

providing, based on the machine learning model, an estimate of the pose of the object for use by the augmented reality presentation device in an augmented reality application associated with the object.

7. The method of claim 6 , wherein:

the augmented reality application is configured to assist a user in performing a particular action with respect to the object; and

the estimate of the pose of the object is used by the augmented reality application to augment the particular input image with integrated guidance information for the performing of the particular action.

8. The method of claim 1 , wherein the accessing of the object image data includes:

receiving a set of preliminary images depicting the object from a plurality of different vantage points; and

generating the object image data based on the set of preliminary images by:

cropping each preliminary image around the object,

manipulating each preliminary image to simulate a straight-on vantage point of the object, and

automatically labeling, or providing a user interface for manually labeling, keypoint features of the object within each preliminary image subsequent to the cropping and manipulating of the preliminary image.

9. The method of claim 1 , wherein:

the plurality of training target images are randomized training target images; and

the particular training target image is generated using one or more random or pseudorandom values to perform:

the selecting of the background image from the set of background images;

the manipulating of the depiction of the object; and

the overlaying of the manipulated depiction of the object onto the selected background image.

10. The method of claim 1 , wherein:

the particular training target image is further generated by manipulating an additional depiction of the object represented within the object image data; and

overlaying the additional manipulated depiction of the object onto the selected background image together with the manipulated depiction of the object.

11. The method of claim 1 , wherein the machine learning model is implemented by a convolutional neural network that includes:

a backbone component configured to progressively process the input images using a series of convolutional layers; and

an anchor-based model head component configured to designate a plurality of anchor areas within the input images and to search each of the plurality of anchor areas for an instance of the object.

12. The method of claim 1 , wherein the machine learning model is implemented by a convolutional neural network that includes:

a backbone component configured to progressively process the input images using a series of convolutional layers; and

a segmentation-based model head component configured to semantically segment the input images to differentiate instances of the object from other image content depicted in the input images.

13. The method of claim 1 , wherein:

the object is a switchboard that includes a plurality of switchboard panels; and

the machine learning model is trained to recognize the object in two phases including:

a first phase in which a prospective recognition of the switchboard is performed based on a subset of keypoint features identified within the input images, and

a second phase in which a confirmed recognition of the switchboard is performed based on the prospective recognition of the first phase and based on respective full sets of keypoint features identified within the input images for each of the plurality of switchboard panels.

14. A system comprising:

a memory storing instructions; and

one or more processors communicatively coupled to the memory and configured to execute the instructions to perform a process comprising:

accessing object image data representative of one or more images depicting an object having a plurality of labeled keypoint features;

generating, based on the object image data, a training target dataset including a plurality of training target images, wherein a particular training target image of the plurality of training target images is generated by:

selecting a background image from a set of background images distinct from the object image data,

manipulating a depiction of the object represented within the object image data, and

overlaying the manipulated depiction of the object onto the selected background image together with an indication of labeled keypoint features for the manipulated depiction; and

training, based on the training target dataset, a machine learning model to recognize and estimate a pose of the object when the object is depicted in input images analyzed using the trained machine learning model.

15. The system of claim 14 , wherein the manipulating of the depiction of the object includes performing at least one of:

a rotation operation to rotate the depiction of the object with respect to two or three spatial dimensions; or

a scaling operation to change a size of the depiction of the object.

16. The system of claim 14 , wherein the manipulating of the depiction of the object includes performing a cropping operation to simulate an occlusion of a portion of the depiction of the object.

17. The system of claim 14 , wherein:

the set of background images corresponds to a background image library configured for use in generating training target datasets for a variety of objects including the object and other objects; and

the background image library comprises background images that are captured independently from image capture associated with the object.

18. The system of claim 14 , wherein:

the particular training target image is further generated by applying, subsequent to the overlaying of the manipulated depiction of the object onto the selected background image, an image processing operation to the particular training target image; and

the image processing operation changes at least one of a contrast attribute, a color attribute, or a saturation attribute of the particular training target image.

19. The system of claim 14 , wherein the process further comprises:

recognizing, based on the machine learning model, the object as depicted within a particular input image captured by an augmented reality presentation device; and

providing, based on the machine learning model, an estimate of the pose of the object for use by the augmented reality presentation device in an augmented reality application associated with the object.

20. A non-transitory computer-readable medium storing instructions that, when executed, direct a processor of a computing device to perform a process comprising:

accessing object image data representative of one or more images depicting an object having a plurality of labeled keypoint features;

generating, based on the object image data, a training target dataset including a plurality of training target images, wherein a particular training target image of the plurality of training target images is generated by:

selecting a background image from a set of background images distinct from the object image data,

manipulating a depiction of the object represented within the object image data, and

overlaying the manipulated depiction of the object onto the selected background image together with an indication of labeled keypoint features for the manipulated depiction; and

training, based on the training target dataset, a machine learning model to recognize and estimate a pose of the object when the object is depicted in input images analyzed using the trained machine learning model.