BACKGROUND
Three dimensional modeling of the human body currently requires large or expensive sensors, such as stereo imaging elements, three-dimensional scanners, depth sensing devices, etc.
DETAILED DESCRIPTION
As is set forth in greater detail below, implementations of the present disclosure are directed to generation of a dimensionally accurate three-dimensional (“3D”) model of a body, such as a human body, based on two-dimensional (“2D”) images of that body, referred to herein as a 2D body image. For example, a user, also referred to herein as a person, may use a 2D camera, such as a digital camera typically included in many of today's portable devices (e.g., cell phones, tablets, laptops, etc.) and obtain a series of 2D body images of their body from different views with respect to the camera. For example, the camera may be stationary, and the user may position themselves in a front view facing toward the camera and the camera may obtain one or more front view images of the body of the user. The user may then turn approximately ninety degrees to the right to a first side view facing the camera, and the camera may obtain one or more first side view images of the body of the user. The user may then turn approximately ninety degrees to the right to a back view facing the camera, and the camera may obtain one or more back view images of the body of the user. The user may then turn approximately ninety degrees to the right to a second side view facing the camera, and the camera may obtain one or more second side view images of the body of the user. As discussed further below, any number of body orientations with respect to the camera and corresponding images thereof may be obtained and used with the disclosed implementations.
Two-dimensional images of the different views of the body of the user may then be processed using one or more processing techniques, as discussed further below, to generate a plurality of predicted body parameters corresponding to the body of the user represented in the images. Those predicted body parameters may then be further processed, as discussed below, to generate a dimensionally accurate 3D model or avatar of the body of the user. In some implementations, the generated 3D model may be further adjusted to match the shape and/or size of the body of the user as represented in one or more of the original 2D body images. Likewise, or in addition thereto, in some implementations, one or more textures (e.g., skin color, clothing, hair, facial features, etc.) may be determined from the original 2D body images and used to augment the 3D model of the user, thereby generating a dimensionally accurate and visually similar 3D model of the user.
The resulting dimensionally accurate 3D model of a body of a user may be utilized for a variety of purposes. For example, it may be presented to a user via an application executing on the same device as used to capture the 2D body images that were processed to generate the 3D model. With the application and the 3D model of the body of the user, the user can track changes to their body, determine body dimensions, measure portions of their body (e.g., arms, waist, shoulders, etc.), determine body mass, body volume, body fat, predict what they would like if they gained muscle, lost fat, etc.
3D modeling of a body from 2D body images begins with the receipt or creation of a 2D body image 102 that includes a representation of the body 103 of the user to be modeled. 2D body images 102 for use with the disclosed implementations may be generated using any conventional imaging element, such as a standard 2D Red, Green, Blue (“RGB”) digital camera that is included on many current portable devices (e.g., tablets, cellular phones, laptops, etc.). The 2D body image may be a still image generated by the imaging element or an image extracted from video generated by the imaging element. Likewise, any number of images may be used with the disclosed implementations.
In some implementations, the user may be instructed to stand in a particular orientation (e.g., front facing toward the imaging element, side facing toward the imaging element, back facing toward the imaging element, etc.) and/or to stand in a particular pose, such as the “A” pose as illustrated in image 102. Likewise, the user may be instructed to stand a distance from the camera such that the body of the user is completely included in a field of view of the imaging element and represented in the generated image 102. Still further, in some implementations, the imaging element may be aligned or positioned at a fixed or stationary point and at a fixed or stationary angle so that images generated by the imaging element are each from the same perspective and encompass the same field of view.
As will be appreciated, a user may elect or opt-in to having a 3D model of the body of the user generated and may also select whether the generated 3D model and/or other information may be used for further training of the disclosed implementations and/or for other purposes.
The 2D body image 102 that includes a representation of the body 103 of the user may then be processed to produce a silhouette 104 of the body 103 of the user represented in the image 102. A variety of techniques may be used to generate the silhouette 104. For example, background subtraction may be used to subtract or black out pixels of the image that correspond to a background of the image while pixels corresponding to the body 103 of user (i.e., foreground) may be assigned a white or other color values. In another example, a semantic segmentation algorithm may be utilized to label background and body (foreground) pixels. For example, a convolutional neural network (“CNN”) may be trained with a semantic segmentation algorithm to determine bodies, such as human bodies, in images.
In some implementations, the silhouette of the body of the user may be normalized in height and centered in the image, as discussed further below. This may be done to further simplify and standardize inputs to a CNN to those on which the CNN was trained. Likewise, a silhouette of the body of the user may be preferred over the representation of the body of the user so that the CNN can focus only on body shape and not skin tone, texture, clothing, etc.
The silhouette 104 of the body may then be processed by one or more other CNNs 106 that are trained to determine body traits, also referred to herein as body features, representative of the body and to produce predicted body parameters that are used to generate a 3D model of the body. In some implementations, the CNN 106 may be trained for multi-mode input to receive as inputs to the CNN the silhouette 104, and one or more known body attributes 105 of the body of the user. For example, a user may provide a height of the body of the user, a weight of the body of the user, a gender of the body of the user, etc., and the CNN may receive one or more of those provided attributes as an input.
Based on the received inputs, the CNN 106 generates predicted body parameters, such as 3D joint locations, body volume, shape of the body, pose angles, etc. In some implementations, the CNN 106 may be trained to predict hundreds of body parameters of the body represented in the image 102.
Utilizing the predicted body parameters, a 3D model of the body is generated. For example, the body parameters may be provided to a body model, such as the Shape Completion and Animation of People (“SCAPE”) body model, a Skinned Multi-Person Linear (“SMPL”) body model, etc., and the body model may generate the 3D model of the body of the user based on those predicted body parameters.
In some implementations, as discussed further below, 3D model refinement 108 may be performed to refine or revise the generated 3D model to better represent the body of the user. For example, the 3D model may be compared to the representation of the body 103 of the user in the image 102 to determine differences between the shape of the body 103 of the user represented in the image 102 and the shape of the 3D model. Based on the determined differences, the silhouette 104 may be revised and the revised silhouette processed by the CNN 106 to produce a revised 3D model of the body of the user. This refinement may continue until there is no or little difference between the shape of the body 103 of the user represented in the image 102 and the shape of the 3D model 110. In other implementations, a 2D model image may be generated from the 3D model and that 2D model image may be compared to the silhouette and/or the 2D body image to determine differences between the 2D model image and the 2D body image or silhouette. Based on the determined differences, the predicted body parameters and/or the 3D model may be revised until the 3D model corresponds to the body of the user represented in the 2D body image and/or the silhouette.
Still further, in some implementations, the 3D model 110 of the body of the user may be augmented with one or more textures, texture augmentation 112, determined from the image 102 of the body of the user. For example, the 3D model may be augmented to have a same or similar color to a skin color of the body 103 represented in the image 102, clothing or clothing colors represented in the image 102 may be used to augment the 3D model, facial features, hair, hair color, etc., of the body of the user represented in the image 102 may be determined and used to augment the 3D model, etc.
The result of the processing illustrated in the transition 100 is a dimensionally accurate 3D model 114 or avatar representative of the body of the user, that has been generated from 2D body images of the body of the user.
In some implementations, multiple 2D body images of a body from different views (e.g., front view, side view, back view, three-quarter view, etc.), such as 2D body images 202-1, 202-2, 202-3, 202-4 through 202-N may be utilized with the disclosed implementations to generate a dimensionally accurate 3D model of the body. In the illustrated example, the first 2D body image 202-1 is an image of a human body 203 oriented in a front view facing a 2D imaging element. The second 2D body image 202-2 is an image of the human body 203 oriented in a first side view facing the 2D imaging element. The third 2D body image 202-3 is an image of the human body 203 oriented in a back view facing the 2D imaging element. The fourth 2D body image 202-4 is an image of the human body 203 oriented in a second side view facing the 2D imaging element. As will be appreciated, any number of 2D body images 202-1 through 202-N may be generated with the view of the human body 203 in any number or orientations with respect to the 2D imaging element.
Each of the 2D body images 202-1 through 202-N are processed to segment pixels of the image that represent the human body from pixels of the image that do not represent the human body to produce a silhouette 204 of the human body as represented in that image. Segmentation may be done through, for example, background subtraction, semantic segmentation, etc. In one example, a baseline image of the background may be known and used to subtract out pixels of the image that correspond to pixels of the baseline image, thereby leaving only foreground pixels that represent the human body. The background pixels may be assigned RGB color values for black (i.e., 0, 0, 0). The remaining pixels may be assigned RGB values for white (i.e., 255, 255, 255) to produce the silhouette 204 or binary segmentation of the human body.
In another example, a CNN utilizing a semantic segmentation algorithm may be trained using images of human bodies, or simulated human bodies to train the CNN to distinguish between pixels that represent human bodies and pixels that do not represent human bodies. In such an example, the CNN may process the image 202 an indicate or label pixels that represent the body (foreground) and pixels that do not represent the body (background). The background pixels may be assigned RGB color values for black (i.e., 0, 0, 0). The remaining pixels may be assigned RGB values for white (i.e., 255, 255, 255) to produce the silhouette or binary segmentation of the human body.
In other implementations, other forms or algorithms, such as edge detection, shape detection, etc., may be used to determine pixels of the image 202 that represent the body and pixels of the image 202 that do not represent the body and a silhouette 204 of the body produced therefrom.
Returning to
In some implementations, in addition to generating a silhouette 204 from the 2D body image, the silhouette may be normalized in size and centered in the image. For example, the silhouette may be cropped by computing a bounding rectangle around the silhouette 204. The silhouette 204 may then be resized according to s, which is a function of a known height h of the user represented in the 2D body image (e.g., the height may be provided by the user):
Where imageh is the input image height, which may be based on the pixels of the image, and μh is the average height of a person (e.g., ˜160 centimeters for females; ˜176 centimeters for males).
Each silhouette 204 representative of the body may then be processed to determine body traits or features of the human body. For example, different CNNs may be trained using silhouettes of bodies, such as human bodies, from different orientations with known features. In some implementations, different CNNs may be trained for different orientations. For example, a first CNN 206A-1 may be trained to determine front view features from front view silhouettes 204-1. A second CNN 206A-2 may be trained to determine right side features from right side silhouettes. A third CNN 206A-3 may be trained to determine back view features from back view silhouettes. A fourth CNN 206A-4 may be trained to determine left side features from left side silhouettes. Different CNNs 206A-1 through 206A-N may be trained for each of the different orientations of silhouettes 204-1 through 204-N. Alternatively, one CNN may be trained to determine features from any orientation silhouette.
In implementations that utilize multiple images of the body 203 to produce multiple sets of features, such as the example illustrated in
Utilizing the predicted body parameters 207, 3D modeling 210 of the body 203 represented in the 2D body images 202 is performed to generate a 3D model of the body 203 represented in the 2D body images 202. For example, the body parameters 207 may be provided to a body model, such as the SCAPE body model, the SMPL body model, etc., and the body model may generate the 3D model of the body 203 represented in the images 202 based on those predicted body parameters 207.
In the illustrated example, 3D model refinement 208 may be performed to refine or revise the generated 3D model to better represent the body 203 represented in the 2D body images 202. For example, the 3D model may be compared to the body 203 represented in one or more of the 2D body images to determine differences between the shape of the body 203 represented in the 2D body image 202 and the shape of the 3D model generated from the predicted body parameters. In some implementations, the 3D model may be compared to a single image, such as image 202-1. In other implementations, the 3D model may be compared to each of the 2D body images 202-1 through 202-N in parallel or sequentially. In still other implementations, one or more 2D model images may be generated from the 3D model and those 2D model images may be compared to the silhouettes and/or the 2D body images to determine differences between the 2D model images and the silhouette/2D body images.
Comparing the 3D model and/or a 2D model image with a 2D body image 202 or silhouette 254 may include determining an approximate pose of the body 203 represented in the 2D body image and adjusting the 3D model to the approximate pose. The 3D model or rendered 2D model image may then be overlaid or otherwise compared to the body 203 represented in the 2D body image 202 and/or represented in the silhouette 204 to determine a difference between the 3D model and the 2D body image.
Based on the determined differences between the 3D model and the body 203 represented in the 2D body image 202, the silhouette 204 generated from that image may be revised to account for those differences. For example, if the 3D model is compared with the body 203 represented in the first image 202-1 and differences are determined, the silhouette 204-1 may be revised based on those differences. Alternatively, the predicted body parameters and/or the 3D model may be revised to account for those differences.
If a silhouette is revised as part of the 3D model refinement 208, the revised silhouette may be processed to determine revised features for the body 203 represented in the 2D body image based on the revised silhouette. The revised features may then be concatenated with the features generated from the other silhouettes or with revised features generated from other revised silhouettes that were produced by the 3D model refinement 208. For example, the 3D model refinement 208 may compare the generated 3D model with the body 203 as represented in two or more 2D body images 202, such as a front image 202-1 and a back image 202-3, differences determined for each of those images, revised silhouettes generated from those differences and revised front view features and revised back view features generated. Those revised features may then be concatenated with the two side view features to produce revised predicted body model parameters. In other implementations, 3D model refinement 208 may compare the 3D model with all views of the body 203 represented in the 2D body images 202 to determine differences and generate revised silhouettes for each of those 2D body images 202-1 through 202-N. Those revised silhouettes may then be processed by the CNNs 206A-1 through 206A-N to produce revised features and those revised features concatenated to produce revised predicted body parameters 207. Finally, the revised predicted body parameters may be processed by 3D modeling 210 to generate a revised 3D model. This process of 3D refinement may continue until there is no or limited difference (e.g., below a threshold difference) between the generated 3D model and the body 203 represented in the 2D body images 202.
In another implementation, 3D model refinement 208 may sequentially compare the 3D model with representations of the body 203 in the different 2D body images 202. For example, 3D model refinement 208 may compare the 3D model with a first representation of the body 203 in a first 2D body image 202-1 to determine differences that are then used to generate a revised silhouette 204-1 corresponding to that first 2D body image 202-1. The revised silhouette may then be processed to produce revised features and those revised features may be concatenated 206B with the features generated from the other silhouettes 204-2 through 204-N to generate revised predicted body parameters, which may be used to generate a revised 3D model. The revised 3D model may then be compared with a next image of the plurality of 2D body images 202 to determine any differences and the process repeated. This process of 3D refinement may continue until there is no or limited difference (e.g., below a threshold difference) between the generated 3D model and the body 203 represented in the 2D body images 202.
In some implementations, upon completion of 3D model refinement 208, the 3D model of the body represented in the 2D body images 202 may be augmented with one or more textures, texture augmentation 212, determined from one or more of the 2D body images 202-1 through 202-N. For example, the 3D model may be augmented to have a same or similar color to a skin color of the body 203 represented the 2D body images 202, clothing or clothing colors represented in the 2D body images 202 may be used to augment the 3D model, facial features, hair, hair color, etc., of the body 203 represented in the 2D body image 202 may be determined and used to augment the 3D model.
Similar to 3D model refinement, the approximate pose of the body in one of the 2D body images 202 may be determined and the 3D model adjusted accordingly so that the texture obtained from that 2D body image 202 may be aligned and used to augment that portion of the 3D model. In some implementations, alignment of the 3D model with the approximate pose of the body 203 may be performed for each 2D body image 202-1 through 202-N so that texture information or data from the different views of the body 203 represented in the different 2D body images 202 may be used to augment the different poses of the resulting 3D model.
The result of the processing illustrated in the transition 200 is a dimensionally accurate 3D model 214 or avatar representative of the body of the user, that has been generated from 2D body images 202 of the body 203 of the user.
In some implementations, multiple 2D body images of a body from different views (e.g., front view, side view, back view, three-quarter view, etc.), such as 2D body images 252-1, 252-2, 252-3, 252-4 through 252-N may be utilized with the disclosed implementations to generate a dimensionally accurate 3D model of the body. In the illustrated example, the first 2D body image 252-1 is an image of a human body 253 oriented in a front view facing a 2D imaging element. The second 2D body image 252-2 is an image of the human body 253 oriented in a first side view facing the 2D imaging element. The third 2D body image 252-3 is an image of the human body 253 oriented in a back view facing the 2D imaging element. The fourth 2D body image 252-4 is an image of the human body 253 oriented in a second side view facing the 2D imaging element. As will be appreciated, any number of 2D body images 252-1 through 252-N may be generated with the view of the human body 253 in any number or orientations with respect to the 2D imaging element.
Each of the 2D body images 252-1 through 252-N are processed to segment pixels of the image that represent the human body from pixels of the image that do not represent the human body to produce a silhouette 254 of the human body as represented in that image. Segmentation may be done through, for example, background subtraction, semantic segmentation, etc. In one example, a baseline image of the background may be known and used to subtract out pixels of the image that correspond to pixels of the baseline image, thereby leaving only foreground pixels that represent the human body. The background pixels may be assigned RGB color values for black (i.e., 0, 0, 0). The remaining pixels may be assigned RGB values for white (i.e., 255, 255, 255) to produce the silhouette 254 or binary segmentation of the human body.
In another example, a CNN utilizing a semantic segmentation algorithm may be trained using images of human bodies, or simulated human bodies to train the CNN to distinguish between pixels that represent human bodies and pixels that do not represent human bodies. In such an example, the CNN may process the image 252 an indicate or label pixels that represent the body (foreground) and pixels that do not represent the body (background). The background pixels may be assigned RGB color values for black (i.e., 0, 0, 0). The remaining pixels may be assigned RGB values for white (i.e., 255, 255, 255) to produce the silhouette or binary segmentation of the human body.
In other implementations, other forms or algorithms, such as edge detection, shape detection, etc., may be used to determine pixels of the image 252 that represent the body and pixels of the image 252 that do not represent the body and a silhouette 254 of the body produced therefrom.
Returning to
Similar to
Where imageh is the input image height, which may be based on the pixels of the image, and μh is the average height of a person (e.g., ˜160 centimeters for females; ˜176 centimeters for males). In some implementations, the body 253 represented in the 2D body image may be similarly resized to correspond to the resized dimensions of the resized silhouette.
Each silhouette 254 representative of the body may then be processed to determine body traits or features of the human body. For example, different CNNs may be trained using silhouettes of bodies, such as human bodies, from different orientations with known features. In some implementations, different CNNs may be trained for different orientations. For example, a first CNN 256A-1 may be trained to determine front view features from front view silhouettes 254-1. A second CNN 256A-2 may be trained to determine right side features from right side silhouettes. A third CNN 256A-3 may be trained to determine back view features from back view silhouettes. A fourth CNN 256A-4 may be trained to determine left side features from left side silhouettes. Different CNNs 256A-1 through 256A-N may be trained for each of the different orientations of silhouettes 254-1 through 254-N. Alternatively, one CNN may be trained to determine features from any orientation silhouette.
In some implementations, the same or different CNNs may also utilize the 2D body image 202 as an input to the CNN that is used to generate and determine the body features. For example, the first CNN 256A-1 may be trained to determine front view features based on inputs of the front view silhouettes 254-1 and/or the 2D body image 252-1. The second CNN 256A-2 may be trained to determine right side features from right side silhouettes and/or the right side 2D body image 252-2. The third CNN 256A-3 may be trained to determine back view features from back view silhouettes and/or the back view 2D body image 252-3. The fourth CNN 256A-4 may be trained to determine left side features from left side silhouettes and/or the left side 2D body image 252-4. Different CNNs 256A-1 through 256A-N may be trained for each of the different orientations of silhouettes 254-1 through 254-N and/or 2D body images 202-1 through 202-N.
In still other implementations, different CNNS may be trained for each of the silhouettes 254 and the 2D body images. For example, the first CNN 256A-1 may be trained to determine front view features from the silhouette 254-1 and another front view CNN may be trained to determine front view features from the 2D body image 252-1. The second CNN 256A-2 may be trained to determine right side view features from the silhouette 254-2 and another right side view CNN may be trained to determine right side view features from the 2D body image 252-2. The third CNN 256A-3 may be trained to determine back view features from the silhouette 254-3 and another back view CNN may be trained to determine back view features from the 2D body image 252-3. The fourth CNN 256A-4 may be trained to determine left side view features from the silhouette 254-4 and another left side view CNN may be trained to determine left side view features from the 2D body image 252-4.
In implementations that utilize multiple images of the body 253 and/or multiple silhouettes to produce multiple sets of features, such as the example illustrated in
Utilizing the predicted body parameters 257, 3D modeling 260 of the body 253 represented in the 2D body images 252 is performed to generate a 3D model of the body 253 represented in the 2D body images 252. For example, the body parameters 257 may be provided to a body model, such as the SCAPE body model, the SMPL body model, etc., and the body model may generate the 3D model of the body 253 represented in the images 252 based on those predicted body parameters 257.
In the illustrated example, 3D model refinement 258 may be performed to refine or revise the generated 3D model to better represent the body 253 represented in the 2D body images 252. For example, as discussed above, the 3D model may be compared to the body 253 represented in one or more of the 2D body images to determine differences between the shape of the body 253 represented in the 2D body image 252 and the shape of the 3D model generated from the predicted body parameters. In some implementations, the 3D model may be compared to a single image, such as image 252-1. In other implementations, the 3D model may be compared to each of the 2D body images 252-1 through 252-N in parallel or sequentially. In still other implementations, one or more 2D model images may be generated from the 3D model and those 2D model images may be compared to the silhouettes and/or the 2D body images to determine differences between the 2D model images and the silhouette/2D body images.
Comparing the 3D model and/or a 2D model image with a 2D body image 252 or silhouette 254 may include determining an approximate pose of the body 253 represented in the 2D body image and adjusting the 3D model to the approximate pose. The 3D model or rendered 2D model image may then be overlaid or otherwise compared to the body 253 represented in the 2D body image 252 and/or represented in the silhouette 254 to determine a difference between the 3D model image and the 2D body image/silhouette.
Based on the determined differences between the 3D model and the body 253 represented in the 2D body image 252, the silhouette 254 generated from that image may be revised to account for those differences. Alternatively, the predicted body parameters and/or the 3D model may be revised to account for those differences.
In some implementations, upon completion of 3D model refinement 258, the 3D model of the body represented in the 2D body images 252 may be augmented with one or more textures, texture augmentation 262, determined from one or more of the 2D body images 252-1 through 252-N. For example, the 3D model may be augmented to have a same or similar color to a skin color of the body 253 represented the 2D body images 252, clothing or clothing colors represented in the 2D body images 252 may be used to augment the 3D model, facial features, hair, hair color, etc., of the body 253 represented in the 2D body image 252 may be determined and used to augment the 3D model.
Similar to 3D model refinement, the approximate pose of the body in one of the 2D body images 252 may be determined and the 3D model adjusted accordingly so that the texture obtained from that 2D body image 252 may be aligned and used to augment that portion of the 3D model. In some implementations, alignment of the 3D model with the approximate pose of the body 253 may be performed for each 2D body image 252-1 through 252-N so that texture information or data from the different views of the body 253 represented in the different 2D body images 252 may be used to augment the different poses of the resulting 3D model.
The result of the processing illustrated in the transition 250 is a dimensionally accurate 3D model 264 or avatar representative of the body of the user, that has been generated from 2D body images 252 of the body 253 of the user.
As discussed above, features or objects expressed in imaging data, such as human bodies, colors, textures or outlines of the features or objects, may be extracted from the data in any number of ways. For example, colors of pixels, or of groups of pixels, in a digital image may be determined and quantified according to one or more standards, e.g., the RGB (“red-green-blue”) color model, in which the portions of red, green or blue in a pixel are expressed in three corresponding numbers ranging from 0 to 255 in value, or a hexadecimal model, in which a color of a pixel is expressed in a six-character code, wherein each of the characters may have a range of sixteen. Moreover, textures or features of objects expressed in a digital image may be identified using one or more computer-based methods, such as by identifying changes in intensities within regions or sectors of the image, or by defining areas of an image corresponding to specific surfaces.
Furthermore, edges, contours, outlines, colors, textures, silhouettes, shapes or other characteristics of objects, or portions of objects, expressed in images may be identified using one or more algorithms or machine-learning tools. The objects or portions of objects may be identified at single, finite periods of time, or over one or more periods or durations. Such algorithms or tools may be directed to recognizing and marking transitions (e.g., the edges, contours, outlines, colors, textures, silhouettes, shapes or other characteristics of objects or portions thereof) within the digital images as closely as possible, and in a manner that minimizes noise and disruptions, and does not create false transitions. Some detection algorithms or techniques that may be utilized in order to recognize characteristics of objects or portions thereof in digital images in accordance with the present disclosure include, but are not limited to, Canny edge detectors or algorithms; Sobel operators, algorithms or filters; Kayyali operators; Roberts edge detection algorithms; Prewitt operators; Frei-Chen methods; semantic segmentation algorithms; background subtraction; or any other algorithms or techniques that may be known to those of ordinary skill in the pertinent arts.
Image processing algorithms, other machine learning algorithms or CNNs may be operated on computer devices of various sizes or types, including but not limited to smartphones or other cell phones, tablets, video cameras or other computer-based machines. Such mobile devices may have limited available computer resources, e.g., network bandwidth, storage capacity or processing power, as compared to larger or more complex computer devices. Therefore, executing computer vision algorithms, other machine learning algorithms, or CNNs on such devices may occupy all or much of the available resources, without any guarantee, or even a reasonable assurance, that the execution of such algorithms will be successful. For example, processing digital 2D body images captured by a user of a portable device (e.g., smartphone, tablet, laptop, webcam) according to one or more algorithms in order to produce a 3D model from the digital images may be an ineffective use of the limited resources that are available on the smartphone or tablet. Accordingly, in some implementations, as discussed herein, some or all of the processing may be performed by one or more computing resources that are remote from the portable device. In some implementations, initial processing of the images to generate binary segmented silhouettes may be performed on the device. Subsequent processing to generate and refine the 3D model may be performed on one or more remote computing resources. For example, the silhouettes may be sent from the portable device to the remote computing resources for further processing. Still further, in some implementations, texture augmentation of the 3D model of the body may be performed on the portable device or remotely.
In some implementations, to increase privacy of the user, only the binary segmented silhouette may be sent from the device for processing on the remote computing resources and the original 2D images that include the representation of the user may be maintained locally on the portable device. In such an example, the rendered 3D model may be sent back to the device and the device may perform texture augmentation of the received 3D model based on those images. Utilizing such a distributed computing arrangement retains user identifiable information on the portable device of the user while at the same time leveraging the increased computing capacity available at remote computing resources.
Machine learning tools, such as artificial neural networks, have been utilized to identify relations between respective elements of apparently unrelated sets of data. An artificial neural network, such as CNN, is a parallel distributed computing processor comprised of individual units that may collectively learn and store experimental knowledge, and make such knowledge available for use in one or more applications. Such a network may simulate the non-linear mental performance of the many neurons of the human brain in multiple layers by acquiring knowledge from an environment through one or more flexible learning processes, determining the strengths of the respective connections between such neurons, and utilizing such strengths when storing acquired knowledge. Like the human brain, an artificial neural network may use any number of neurons in any number of layers, including an input layer, an output layer, and one or more intervening hidden layers. In view of their versatility, and their inherent mimicking of the human brain, machine learning tools including not only artificial neural networks but also nearest neighbor methods or analyses, factorization methods or techniques, K-means clustering analyses or techniques, similarity measures such as log likelihood similarities or cosine similarities, latent Dirichlet allocations or other topic models, or latent semantic analyses have been utilized in image processing applications.
Artificial neural networks may be trained to map inputted data to desired outputs by adjusting the strengths of the connections between one or more neurons, which are sometimes called synaptic weights. An artificial neural network may have any number of layers, including an input layer, an output layer, and any number of intervening hidden layers. Each of the neurons in a layer within a neural network may receive one or more inputs and generate one or more outputs in accordance with an activation or energy function, with parameters corresponding to the various strengths or synaptic weights. Likewise, each of the neurons within a network may be understood to have different activation or energy functions; in this regard, such a network may be dubbed a heterogeneous neural network. In some neural networks, at least one of the activation or energy functions may take the form of a sigmoid function, wherein an output thereof may have a range of zero to one or 0 to 1. In other neural networks, at least one of the activation or energy functions may take the form of a hyperbolic tangent function, wherein an output thereof may have a range of negative one to positive one, or −1 to +1. Thus, the training of a neural network according to an identity function results in the redefinition or adjustment of the strengths or weights of such connections between neurons in the various layers of the neural network, in order to provide an output that most closely approximates or associates with the input to the maximum practicable extent.
Artificial neural networks may typically be characterized as either feedforward neural networks or recurrent neural networks, and may be fully or partially connected. In a feedforward neural network, e.g., a convolutional neural network, information specifically flows in one direction from an input layer to an output layer, while in a recurrent neural network, at least one feedback loop returns information regarding the difference between the actual output and the targeted output for training purposes. Additionally, in a fully connected neural network architecture, each of the neurons in one of the layers is connected to all of the neurons in a subsequent layer. By contrast, in a sparsely connected neural network architecture, the number of activations of each of the neurons is limited, such as by a sparsity parameter.
Moreover, the training of a neural network is typically characterized as supervised or unsupervised. In supervised learning, a training set comprises at least one input and at least one target output for the input. Thus, the neural network is trained to identify the target output, to within an acceptable level of error. In unsupervised learning of an identity function, such as that which is typically performed by a sparse autoencoder, target output of the training set is the input, and the neural network is trained to recognize the input as such. Sparse autoencoders employ backpropagation in order to train the autoencoders to recognize an approximation of an identity function for an input, or to otherwise approximate the input. Such backpropagation algorithms may operate according to methods of steepest descent, conjugate gradient methods, or other like methods or techniques, in accordance with the systems and methods of the present disclosure. Those of ordinary skill in the pertinent art would recognize that any algorithm or method may be used to train one or more layers of a neural network. Likewise, any algorithm or method may be used to determine and minimize the error in an output of such a network. Additionally, those of ordinary skill in the pertinent art would further recognize that the various layers of a neural network may be trained collectively, such as in a sparse autoencoder, or individually, such that each output from one hidden layer of the neural network acts as an input to a subsequent hidden layer.
Once a neural network has been trained to recognize dominant characteristics of an input of a training set, e.g., to associate an image with a label, a category, a cluster or a pseudolabel thereof, to within an acceptable tolerance, an input and/or multiple inputs, in the form of an image, silhouette, features, known traits corresponding to the image, etc., may be provided to the trained network, and an output generated therefrom. For example, the CNN discussed above may receive as inputs a generated silhouette and one or more body attributes (e.g., height, weight, gender) corresponding to the body represented by the silhouette. The trained CNN may then produce as outputs the predicted features corresponding to those inputs.
Referring to
The system 300 of
The body composition system 310 of
The imaging element 320 may comprise any form of optical recording sensor or device that may be used to photograph or otherwise record information or data regarding a body of the user, or for any other purpose. As is shown in
The portable device 330 may be used in any location and any environment to generate 2D body images that represent a body of the user. In some implementations, the portable device may be positioned such that it is stationary and approximately vertical (within approximately ten-degrees of vertical) and the user may position their body within a field of view of the imaging element 320 of the portable device at different orientations so that the imaging element 320 of the portable device may generate 2D body images that include a representation of the body of the user from different orientations.
The portable device 330 may also include one or more applications 323 stored in memory that may be executed by the processor 326 of the portable device to cause the processor of the portable device to perform various functions or actions. For example, when executed, the application 323 may provide instructions to a user regarding placement of the portable device, positioning of the body of the user within the field of view of the imaging element 320 of the portable device, orientation of the body of the user, etc. Likewise, in some implementations, the application may present a 3D model, generated from the 2D body images in accordance with the described implementations, to the user and allow the user to interact with the 3D model. For example, a user may rotate the 3D model to view different angles of the 3D model, obtain approximately accurate measurements of the body of the user from the dimensions of the 3D model, view body fat, body mass, volume, etc. Likewise, in some implementations, the 3D model may be modified by request of the user to simulate what the body of the user may look like under certain conditions, such as loss of weight, gain of muscle, etc.
The external media storage facility 370 may be any facility, station or location having the ability or capacity to receive and store information or data, such as silhouettes, simulated or rendered 3D models of bodies, textures, body dimensions, etc., received from the body composition system 310, and/or from the portable device 330. As is shown in
The network 380 may be any wired network, wireless network, or combination thereof, and may comprise the Internet in whole or in part. In addition, the network 380 may be a personal area network, local area network, wide area network, cable network, satellite network, cellular telephone network, or combination thereof. The network 380 may also be a publicly accessible network of linked networks, possibly operated by various distinct parties, such as the Internet. In some implementations, the network 380 may be a private or semi-private network, such as a corporate or university intranet. The network 380 may include one or more wireless networks, such as a Global System for Mobile Communications (GSM) network, a Code Division Multiple Access (CDMA) network, a Long Term Evolution (LTE) network, or some other type of wireless network. Protocols and components for communicating via the Internet or any of the other aforementioned types of communication networks are well known to those skilled in the art of computer communications and thus, need not be described in more detail herein.
The computers, servers, devices and the like described herein have the necessary electronics, software, memory, storage, databases, firmware, logic/state machines, microprocessors, communication links, displays or other visual or audio user interfaces, printing devices, and any other input/output interfaces to provide any of the functions or services described herein and/or achieve the results described herein. Also, those of ordinary skill in the pertinent art will recognize that users of such computers, servers, devices and the like may operate a keyboard, keypad, mouse, stylus, touch screen, or other device (not shown) or method to interact with the computers, servers, devices and the like, or to “select” an item, link, node, hub or any other aspect of the present disclosure.
The body composition system 310, the portable device 330 or the external media storage facility 370 may use any web-enabled or Internet applications or features, or any other client-server applications or features including E-mail or other messaging techniques, to connect to the network 380, or to communicate with one another, such as through short or multimedia messaging service (SMS or MMS) text messages. For example, the servers 312-1, 312-2 . . . 312-M may be adapted to transmit information or data in the form of synchronous or asynchronous messages from the body composition system 310 to the processor 326 or other components of the portable device 330, or any other computer device in real time or in near-real time, or in one or more offline processes, via the network 380. Those of ordinary skill in the pertinent art would recognize that the body composition system 310, the portable device 330 or the external media storage facility 370 may operate any of a number of computing devices that are capable of communicating over the network, including but not limited to set-top boxes, personal digital assistants, digital media players, web pads, laptop computers, desktop computers, electronic book readers, cellular phones, and the like. The protocols and components for providing communication between such devices are well known to those skilled in the art of computer communications and need not be described in more detail herein.
The data and/or computer executable instructions, programs, firmware, software and the like (also referred to herein as “computer executable” components) described herein may be stored on a computer-readable medium that is within or accessible by computers or computer components such as the servers 312-1, 312-2 . . . 312-M, the processor 326, the servers 372-1, 372-2 . . . 372-J, or any other computers or control systems utilized by the body composition system 310, the portable device 330, applications 323, or the external media storage facility 370, and having sequences of instructions which, when executed by a processor (e.g., a central processing unit, or “CPU”), cause the processor to perform all or a portion of the functions, services and/or methods described herein. Such computer executable instructions, programs, software and the like may be loaded into the memory of one or more computers using a drive mechanism associated with the computer readable medium, such as a floppy drive, CD-ROM drive, DVD-ROM drive, network interface, or the like, or via external connections.
Some implementations of the systems and methods of the present disclosure may also be provided as a computer-executable program product including a non-transitory machine-readable storage medium having stored thereon instructions (in compressed or uncompressed form) that may be used to program a computer (or other electronic device) to perform processes or methods described herein. The machine-readable storage media of the present disclosure may include, but is not limited to, hard drives, floppy diskettes, optical disks, CD-ROMs, DVDs, ROMs, RAMs, erasable programmable ROMs (“EPROM”), electrically erasable programmable ROMs (“EEPROM”), flash memory, magnetic or optical cards, solid-state memory devices, or other types of media/machine-readable medium that may be suitable for storing electronic instructions. Further, implementations may also be provided as a computer executable program product that includes a transitory machine-readable signal (in compressed or uncompressed form). Examples of machine-readable signals, whether modulated using a carrier or not, may include, but are not limited to, signals that a computer system or machine hosting or running a computer program can be configured to access, or including signals that may be downloaded through the Internet or other networks.
Utilizing binary silhouettes 404 of bodies improves the accuracy of the feature determination CNN 406A as it can focus purely on size, shape, dimensions, etc. of the body, devoid of any other aspects (e.g., color, clothing, hair, etc.). In other implementations, the use of the 2D body images in conjunction with or independent of the silhouettes provides additional data, such as shadows, skin tone, etc., that the feature determination CNN 406A may utilize in determining and producing a set of features representative of the body represented in the 2D body image.
The features output from the feature determination CNNs 406A, in the disclosed implementation, are received as inputs to the concatenation CNN 406B. Likewise, in some implementations, the concatenation CNN 406B may be trained to receive other inputs, such as body attributes 405.
As discussed, the concatenation CNN 406B may be trained to receive the inputs of features, and optionally other inputs, produce concatenated features, and produce as outputs a set of predicted body parameters 407 corresponding to the body represented in the 2D body images. In some implementations, the predicted body parameters may include hundreds of parameters, including, but not limited to, shape, pose, volume, joint position, etc., of the represented body.
The example process 500 begins upon receipt of one or more 2D body images of a body, as in 502. As noted above, the disclosed implementations are operable with any number of 2D body images for use in generating a 3D model of that body. For example, in some implementations, a single 2D body image may be used. In other implementations, two, three, four, or more 2D body images may be used.
As discussed above, the 2D body images may be generated using any 2D imaging element, such as a camera on a portable device, a webcam, etc. The received 2D body images are then segmented to produce a binary silhouette of the body represented in the one or more 2D body images, as in 504. As discussed above, one or more segmentation techniques, such as background subtraction, semantic segmentation, Canny edge detectors or algorithms, Sobel operators, algorithms or filters, Kayyali operators, Roberts edge detection algorithms, Prewitt operators, Frei-Chen methods, or any other algorithms or techniques that may be known to those of ordinary skill in the pertinent arts.
In addition, in some implementations, the silhouettes may be normalized in height and centered in the image before further processing, as in 506. For example, the silhouettes may be normalized to a standard height based on a function of a known or provided height of the body of the user represented in the image and an average height (e.g., average height of female body, average height of male body). In some implementations, the average height may be more specific than just gender. For example, the average height may be the average height of a gender and a race corresponding to the body, or a gender and a location (e.g., United States) of the user, etc.
The normalized and centered silhouette may then be processed by one or more neural networks, such as one or more CNNs as discussed above, to generate predicted body parameters representative of the body represented in the 2D body images, as in 508. As discussed above, there may be multiple steps involved in body parameter prediction. For example, each silhouette may be processed using CNNs trained for the respective orientation of the silhouette to generate sets of features of the body as determined from the silhouette. The sets of features generated from the different silhouette may then be processed using a neural network, such as a CNN, to concatenate the features and generate the predicted body parameters representative of the body represented in the 2D body images.
The predicted body parameters may then be provided to one or more body models, such as an SMPL body model or a SCAPE body model and the body model may generate a 3D model for the body represented in the 2D body images, as in 510. In addition, in some implementations, the 3D model may be revised, if necessary, to more closely correspond to the actual image of the body of the user, as in 600. 3D model refinement is discussed above, and discussed further below with respect
As discussed below, the 3D model refinement process 600 (
After adjustment of the silhouette and generation of a 3D model from adjusted body parameters, or after receipt of the adjusted 3D model from
The 3D model of the body of the user may then be adjusted to correspond to the determined pose of the body in the 2D body image, as in 604. With the 3D model adjusted to approximately the same pose as the user represented in the image, the shape of the 3D model may be compared to the shape of the body in the 2D body image and/or the silhouette to determine any differences between the 3D model and the representation of the body in the 2D body image and/or silhouette, as in 606.
In some implementations, it may be determined whether any determined difference is above a minimum threshold (e.g., 2%). If it is determined that there is a difference between the 3D model and the body represented in one or more of the 2D body images, the silhouette may be adjusted. The silhouette may then be used to generate revised body parameters for the body represented in the 2D body images, as discussed above with respect to
The 3D model of the body of the user may then be adjusted to correspond to the determined pose of the body in the 2D body image, as in 654. With the 3D model adjusted to approximately the same pose as the user represented in the image, a 2D model image from the 3D model is generated, as in 656. The 2D model image may be generated, for example, by converting or imaging the 3D model into a 2D image with the determined pose, as if a digital 2D image of the 3D model had been generated. Likewise, the 2D model image may be a binary image with pixels corresponding to the model having a first set of values (e.g., white—RGB values of 255, 255, 255) and pixels that do not represent the model having a second set of values (e.g., black—RGB values of 0, 0, 0)
The 2D model image is then compared with the 2D body image and/or the silhouette to determine any differences between the 2D model image and the representation of the body in the 2D body image and/or silhouette, as in 658. For example, the 2D model image may be aligned with the 2D body image and/or the silhouette and pixels between the images compared to determine differences between the pixels values. In implementations in which the pixels are binary (e.g., white or black) an error (e.g., % difference) may be determined as a difference in pixel values between the 2D model image and the 2D body image. That error is differentiable and may be utilized to adjust the predicted body parameters and, as a result, the shape of the 3D model.
In some implementations, it may be determined whether any determined difference is above a minimum threshold (e.g., 2%). If it is determined that there is a difference between the 2D model image and the body represented in one or more of the 2D body images/silhouette, the 3D model and/or the predicted body parameters may be adjusted to correspond to the shape and/or size of body represented in the 2D body image and/or the silhouette, as in 660. This example process 650 may continue until there is no difference between the 2D model image and the 2D body image/silhouette, or the difference is below a minimum threshold. As discussed above, the revised 3D model produced from the example process 650, or if no adjustment are necessary, the 3D model is returned to example process 500 at block 512 and the process 500 continues.
Although the disclosure has been described herein using exemplary techniques, components, and/or processes for implementing the systems and methods of the present disclosure, it should be understood by those skilled in the art that other techniques, components, and/or processes or other combinations and sequences of the techniques, components, and/or processes described herein may be used or performed that achieve the same function(s) and/or result(s) described herein and which are included within the scope of the present disclosure.
Additionally, in accordance with the present disclosure, the training of machine learning tools (e.g., artificial neural networks or other classifiers) and the use of the trained machine learning tools to generate 3D models of a body based on one or more 2D images of that body may occur on multiple, distributed computing devices, or on a single computing device.
Furthermore, although some implementations of the present disclosure reference the use of separate machine learning tools or separate CNNs for processing silhouettes, for concatenating features determined from silhouettes, and/or for generating the 3D model of the body, the systems and methods of the present disclosure are not so limited. Features, predicated body parameters, and/or 3D models may be determined and generated using a single CNN, or with two or more CNNs, in accordance with the present disclosure.
Likewise, while the above discussions focus primarily on generating a 3D model of a body using multiple 2D body images of the body, in some implementations, the 3D model may be generated based on a single 2D body image of the body. In other implementations, two or more 2D body images of the body from different orientations may be used with the disclosed implementations. In general, the more 2D body images of the body from different orientations, the more accurate the final representation and dimensions of the 3D model may be.
Still further, while the above implementations are described with respect generating 3D models of human bodies represented in 2D body images, in other implementations, non-human bodies, such as dogs, cats, or other animals may be modeled in 3D based on 2D representations of those bodies. Accordingly, the use of human body in the disclosed implementations should not be considered limiting.
It should be understood that, unless otherwise explicitly or implicitly indicated herein, any of the features, characteristics, alternatives or modifications described regarding a particular implementation herein may also be applied, used, or incorporated with any other implementation described herein, and that the drawings and detailed description of the present disclosure are intended to cover all modifications, equivalents and alternatives to the various implementations as defined by the appended claims. Moreover, with respect to the one or more methods or processes of the present disclosure described herein, including but not limited to the flow charts shown in
Conditional language, such as, among others, “can,” “could,” “might,” or “may,” unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey in a permissive manner that certain implementations could include, or have the potential to include, but do not mandate or require, certain features, elements and/or steps. In a similar manner, terms such as “include,” “including” and “includes” are generally intended to mean “including, but not limited to.” Thus, such conditional language is not generally intended to imply that features, elements and/or steps are in any way required for one or more implementations or that one or more implementations necessarily include logic for deciding, with or without user input or prompting, whether these features, elements and/or steps are included or are to be performed in any particular implementation.
The elements of a method, process, or algorithm described in connection with the implementations disclosed herein can be embodied directly in hardware, in a software module stored in one or more memory devices and executed by one or more processors, or in a combination of the two. A software module can reside in RAM, flash memory, ROM, EPROM, EEPROM, registers, a hard disk, a removable disk, a CD-ROM, a DVD-ROM or any other form of non-transitory computer-readable storage medium, media, or physical computer storage known in the art. An example storage medium can be coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The storage medium can be volatile or nonvolatile. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.
Disjunctive language such as the phrase “at least one of X, Y, or Z,” or “at least one of X, Y and Z,” unless specifically stated otherwise, is otherwise understood with the context as used in general to present that an item, term, etc., may be either X, Y, or Z, or any combination thereof (e.g., X, Y, and/or Z). Thus, such disjunctive language is not generally intended to, and should not, imply that certain implementations require at least one of X, at least one of Y, or at least one of Z to each be present.
Unless otherwise explicitly stated, articles such as “a” or “an” should generally be interpreted to include one or more described items. Accordingly, phrases such as “a device configured to” are intended to include one or more recited devices. Such one or more recited devices can also be collectively configured to carry out the stated recitations. For example, “a processor configured to carry out recitations A, B and C” can include a first processor configured to carry out recitation A working in conjunction with a second processor configured to carry out recitations B and C.
Language of degree used herein, such as the terms “about,” “approximately,” “generally,” “nearly” or “substantially” as used herein, represent a value, amount, or characteristic close to the stated value, amount, or characteristic that still performs a desired function or achieves a desired result. For example, the terms “about,” “approximately,” “generally,” “nearly” or “substantially” may refer to an amount that is within less than 10% of, within less than 5% of, within less than 1% of, within less than 0.1% of, and within less than 0.01% of the stated amount.
Although the invention has been described and illustrated with respect to illustrative implementations thereof, the foregoing and various other additions and omissions may be made therein and thereto without departing from the spirit and scope of the present disclosure.