Abstract: In some embodiments, a method trains a first parameter of a differentiable proxy codec to encode source content based on a first loss between first compressed source content and second compressed source content that is output by a target codec. A pre-processor pre-processes a source image to output a pre-processed source image, the pre-processing being based on a second parameter. The differentiable proxy codec encodes the pre-processed source image into a compressed pre-processed source image based on the first parameter. The method determines a second loss between the source image and the compressed pre-processed source image and determines an adjustment to the first parameter based on the second loss. The adjustment is used to adjust the second parameter of the pre-processor based on the second loss.

Abstract: A system includes a hardware processor and a memory storing software code. The hardware processor executes the software code to receive a three-dimensional (3-D) image including a plurality of pixels each having a plurality of bins with respective depth values, select, for a first bin of the plurality of bins of a first pixel, a second bin in each of one or more nearest neighbor pixels of the first pixel, the second bin in each of the one or more nearest neighbor pixels having a most similar depth value to a depth value of the first bin. The hardware processor further executes the software code to generate a first depth-aware bin group including the first bin and the second bin in each of the one or more nearest neighbor pixels, and process, using the first depth-aware bin group, the 3-D image to produce a corresponding 3-D output image.

Abstract: A system includes a pre-processor configured to receive three-dimensional (3-D) image data, flatten the 3-D image data to produce corresponding two-dimensional (2-D) image data, and concatenate the 3-D image data and the corresponding 2-D image data to provide concatenated image data. The system further includes an encoder including one or more first neural networks (NNs), the encoder configured to use the one or more first NNs to encode the concatenated image data to provide encoded data, a decoder including one or more second NNs, the decoder configured to use the one or more second NNs to decode the encoded data to provide decoded data, and a reconstructor including a plurality of hybrid 2-D/3-D reconstructors configured to reconstruct the decoded data to provide a denoised 3-D output image corresponding to the 3-D image data.

Abstract: Techniques are disclosed for training and applying a denoising model. The denoising model includes multiple specialized denoisers and a generalizer, each of which is a machine learning model. The specialized denoisers are trained to denoise images associated with specific ranges of noise parameters. The generalizer is trained to generate per-pixel denoising kernels for denoising images associated with arbitrary noise parameters using outputs of the specialized denoisers. Subsequent to training, a noisy image, such as a live-action image or a rendered image, can be denoised by inputting the noisy image into the specialized denoisers to obtain intermediate denoised images that are then input, along with the noisy image, into the generalizer to obtain per-pixel denoising kernels, which can be normalized and applied to denoise the noisy image.

Abstract: The disclosed matting technique comprises receiving a video feed comprising a plurality of temporally ordered video frames as the video frames are captured by a video capture device, generating, using one or more machine learning models, an image mask corresponding to each video frame included in the video feed and a depth estimate corresponding to each video frame included in the video feed, and, for each video frame in the video feed, transmitting, in real-time, the video frame, the corresponding image mask, and the corresponding depth estimate to a compositing system. The compositing system composites a computer-generated element with the video frame based on the corresponding image mask and the corresponding depth estimate.

Abstract: A system includes a hardware processor, a system memory storing a software code, and a machine learning (ML) model trained using style loss to predict image noise. The hardware processor is configured to execute the software code to receive a clean image and at least one noise setting of a camera used to capture a version of the clean image that includes noise, and provide the clean image and the at least one noise setting as a noise generation input to the ML model. The hardware processor is further configured to execute the software code to generate, using the ML model and based on the noise generation input, a synthesized noise map for renoising the clean image.

Abstract: The exemplary embodiments relate to converting Standard Dynamic Range (SDR) content to High Dynamic Range (HDR) content using a machine learning system. In some embodiments, the neural network is trained to convert an input SDR image into an HDR image using the encoded representation of a training SDR image and a training HDR image. In other embodiments, the neural network is trained to convert an input SDR image into an HDR image using a predefined set of color grading actions and the training images.

Abstract: The exemplary embodiments relate to converting Standard Dynamic Range (SDR) content to High Dynamic Range (HDR) content using a machine learning system. In some embodiments, the neural network is trained to convert an input SDR image into an HDR image using the encoded representation of a training SDR image and a training HDR image. In other embodiments, the neural network is trained to convert an input SDR image into an HDR image using a predefined set of color grading actions and the training images.

Abstract: Techniques are disclosed for denoising videos. In some embodiments, video frames are denoised using a denoising model that includes an encoder-decoder architecture and attention modules. During training of the denoising model, the attention modules learn weightings to upweight certain dimensions of input features to help pixel registration, remove ghosting artifacts, and improve temporal consistency when the frames of a video are being denoised. The denoising model can also be used to train a student denoising model that has a same architecture as, but is smaller and faster than, the denoising model. After training, noisy video frames can be input into the denoising model and/or the student denoising model to generate corresponding denoised video frames.

Abstract: Techniques are disclosed for enhancing videos using a machine learning model that is a temporally-consistent transformer model. The machine learning model processes blocks of frames of a video in which the temporally first input video frame of each block of frames is a temporally second to last output video frame of a previous block of frames. After the machine learning model is trained, blocks of video frames, or features extracted from the video frames, can be warped using an optical flow technique and transformed using a wavelet transform technique. The transformed video frames are concatenated along a channel dimension and input into the machine learning model that generates corresponding processed video frames.

Abstract: A method of generating a training data set for training an image matting machine learning model includes receiving a plurality of foreground images, generating a plurality of composited foreground images by compositing randomly selected foreground images from the plurality of foreground images, and generating a plurality of training images by compositing each composited foreground image with a randomly selected background image. The training data set includes the plurality of training images.

Abstract: The exemplary embodiments relate to converting Standard Dynamic Range (SDR) content to High Dynamic Range (HDR) content using a machine learning system. In some embodiments, the neural network is trained to convert an input SDR image into an HDR image using the encoded representation of a training SDR image and a training HDR image. In other embodiments, the neural network is trained to convert an input SDR image into an HDR image using a predefined set of color grading actions and the training images.

Abstract: Novel systems and methods are described for creating, compressing, and distributing video or image content graded for a plurality of displays with different dynamic ranges. In implementations, the created content is “continuous dynamic range” (CDR) content—a novel representation of pixel-luminance as a function of display dynamic range. The creation of the CDR content includes grading a source content for a minimum dynamic range and a maximum dynamic range, and defining a luminance of each pixel of an image or video frame of the source content as a continuous function between the minimum and the maximum dynamic ranges. In additional implementations, a novel graphical user interface for creating and editing the CDR content is described.

Abstract: A method of generating a training data set for training an image matting machine learning model includes receiving a plurality of foreground images, generating a plurality of com posited foreground images by com positing randomly selected foreground images from the plurality of foreground images, and generating a plurality of training images by compositing each composited foreground image with a randomly selected background image. The training data set includes the plurality of training images.

Abstract: Techniques are disclosed for training and applying a denoising model. The denoising model includes multiple specialized denoisers and a generalizer, each of which is a machine learning model. The specialized denoisers are trained to denoise images associated with specific ranges of noise parameters. The generalizer is trained to generate per-pixel denoising kernels for denoising images associated with arbitrary noise parameters using outputs of the specialized denoisers. Subsequent to training, a noisy image, such as a live-action image or a rendered image, can be denoised by inputting the noisy image into the specialized denoisers to obtain intermediate denoised images that are then input, along with the noisy image, into the generalizer to obtain per-pixel denoising kernels, which can be normalized and applied to denoise the noisy image.

Abstract: Techniques are disclosed for image matting. In particular, embodiments decompose the matting problem of estimating foreground opacity into the targeted subproblems of estimating a background using a first trained neural network, estimating a foreground using a second neural network and the estimated background as one of the inputs into the second neural network, and estimating an alpha matte using a third neural network and the estimated background and foreground as two of the inputs into the third neural network. Such a decomposition is in contrast to traditional sampling-based matting approaches that estimated foreground and background color pairs together directly for each pixel. By decomposing the matting problem into subproblems that are easier for a neural network to learn compared to traditional data-driven techniques for image matting, embodiments disclosed herein can produce better opacity estimates than such data-driven techniques as well as sampling-based and affinity-based matting approaches.

Abstract: Techniques are disclosed for image matting. In particular, embodiments decompose the matting problem of estimating foreground opacity into the targeted subproblems of estimating a background using a first trained neural network, estimating a foreground using a second neural network and the estimated background as one of the inputs into the second neural network, and estimating an alpha matte using a third neural network and the estimated background and foreground as two of the inputs into the third neural network. Such a decomposition is in contrast to traditional sampling-based matting approaches that estimated foreground and background color pairs together directly for each pixel. By decomposing the matting problem into subproblems that are easier for a neural network to learn compared to traditional data-driven techniques for image matting, embodiments disclosed herein can produce better opacity estimates than such data-driven techniques as well as sampling-based and affinity-based matting approaches.

Abstract: Systems and methods are disclosed for weighting the image quality prediction of any visual-attention-agnostic quality metric with a saliency map. By accounting for the salient regions of an image or video frame, the disclosed systems and methods may dramatically improve the precision of the visual-attention-agnostic quality metric during image or video quality assessment. In one implementation, a method of saliency-weighted video quality assessment includes: determining a per-pixel image quality vector of an encoded video frame; determining per-pixel saliency values of the encoded video frame or a reference video frame corresponding to the encoded video frame; and computing a saliency-weighted image quality metric of the encoded video frame by weighting the per-pixel image quality vector using the per-pixel saliency values.

Abstract: Embodiments can provide a strategy for controlling information flow both from known opacity regions to unknown regions, as well as within the unknown region itself. This strategy is formulated through the use and refinement of various affinity definitions. As a result of this strategy, a final linear system can be obtained, which can be solved in closed form. One embodiment pertains to identifying opacity information flows. The opacity information flow may include one or more of flows from pixels in the image that have similar colors to a target pixel, flows from pixels in the foreground and background to the target pixel, flows from pixels in the unknown opacity region in the image to the target pixel, flows from pixels immediately surrounding the target pixels in the image to the target pixel, and any other flow.

Abstract: The disclosure provides an approach for edge-aware spatio-temporal filtering. In one embodiment, a filtering application receives as input a guiding video sequence and video sequence(s) from additional channel(s). The filtering application estimates a sparse optical flow from the guiding video sequence using a novel binary feature descriptor integrated into the Coarse-to-fine PatchMatch method to compute a quasi-dense nearest neighbor field. The filtering application then performs spatial edge-aware filtering of the sparse optical flow (to obtain a dense flow) and the additional channel(s), using an efficient evaluation of the permeability filter with only two scan-line passes per iteration. Further, the filtering application performs temporal filtering of the optical flow using an infinite impulse response filter that only requires one filter state updated based on new guiding video sequence video frames.