Abstract: Methods, systems, apparatuses, and computer-readable storage mediums described herein are directed to techniques for efficient data encoding for neural network training. In particular, the embodiments described herein train a DNN based on a selective encoding (e.g., compressing) of data structures that are generated during training. For example, multiple training sessions may be performed where, in each training session, a different set of data structures performed by various operators of the DNN are encoded. Memory allocation information generated based on each training session is analyzed to determine which combination of encoded data structures results in a reduction of memory required to train the DNN.

Abstract: Embodiments of the present disclosure include systems and methods for training masked language models based on partial sequences of tokens. A sequence of tokens for training a transformer model is received. A defined proportion of the sequence of tokens is selected. Each value of the defined proportion of the sequence of tokens is replaced with a defined value.

Abstract: Techniques related to synthesizing an image of a person in an unseen pose are discussed. Such techniques include detecting a body part occlusion for a body part in a representation of the person in a first image and, in response to the detected occlusion, projecting a representation of the body part from a second image having a different view into the first image. A geometric transformation based on a source pose of the person and a target pose is then applied to the merged image to generate a synthesized image comprising a representation of the person in the target pose.

Abstract: A mechanism is described for facilitating smart point cloud reconstruction of objects in visual scenes in computing environments. An apparatus of embodiments, as described herein, includes one or more processors including one or more graphics processors, and photo-consistency logic to perform line searches on cloud points of an object to enhance photo-consistency between multiple camera views associated with a visual hull encompassing the object in a scene captured by multiple cameras coupled to the one or more processors. The apparatus may further include refinement and application logic to perform a final reconstruction of the object based on the enhanced photo-consistency, where the scene having the reconstructed object is displayed at a display device.

Abstract: Techniques related to synthesizing an image of a person in an unseen pose are discussed. Such techniques include detecting a body part occlusion for a body part in a representation of the person in a first image and, in response to the detected occlusion, projecting a representation of the body part from a second image having a different view into the first image. A geometric transformation based on a source pose of the person and a target pose is then applied to the merged image to generate a synthesized image comprising a representation of the person in the target pose.

Abstract: Techniques related to synthesizing an image of a person in an unseen pose are discussed. Such techniques include detecting a body part occlusion for a body part in a representation of the person in a first image and, in response to the detected occlusion, projecting a representation of the body part from a second image having a different view into the first image. A geometric transformation based on a source pose of the person and a target pose is then applied to the merged image to generate a synthesized image comprising a representation of the person in the target pose.

Abstract: An apparatus to facilitate three dimensional (3D) position estimation is disclosed. The apparatus includes one or more processors to receive a plurality of images captured by a camera array during a live event, locate key-points of human joints of a plurality of event participants included in the images, associate key-points of each participant across the images and recover a 3D body position of each of the plurality of participants based on the associated key-points.

Abstract: A mechanism is described for facilitating smart point cloud reconstruction of objects in visual scenes in computing environments. An apparatus of embodiments, as described herein, includes one or more processors including one or more graphics processors, and photo-consistency logic to perform line searches on cloud points of an object to enhance photo-consistency between multiple camera views associated with a visual hull encompassing the object in a scene captured by multiple cameras coupled to the one or more processors. The apparatus may further include refinement and application logic to perform a final reconstruction of the object based on the enhanced photo-consistency, where the scene having the reconstructed object is displayed at a display device.

Abstract: Techniques related to synthesizing an image of a person in an unseen pose are discussed. Such techniques include detecting a body part occlusion for a body part in a representation of the person in a first image and, in response to the detected occlusion, projecting a representation of the body part from a second image having a different view into the first image. A geometric transformation based on a source pose of the person and a target pose is then applied to the merged image to generate a synthesized image comprising a representation of the person in the target pose.

Abstract: An apparatus to facilitate three dimensional (3D) position estimation is disclosed. The apparatus includes one or more processors to receive a plurality of images captured by a camera array during a live event, locate key-points of human joints of a plurality of event participants included in the images, associate key-points of each participant across the images and recover a 3D body position of each of the plurality of participants based on the associated key-points.