Abstract: In particular embodiments, a computing system may access a plurality of images for pre-training a first machine-learning model that includes an encoder and a decoder. Using each image, the system may pre-train the model by dividing the image into a set a patches, selecting a first subset of the patches to be visible and a second subset of the patches to be masked during the pre-training, processing, using the encoder, the first subset of patches to generate corresponding first latent representations, processing, using the decoder, the first latent representations corresponding to the first subset of patches and mask tokens corresponding to the second subset of patches to generate reconstructed patches corresponding to the second subset of patches, the reconstructed patches and the first subset of patches being used to generate a reconstructed image, and updating the model based on comparisons between the image and the reconstructed image.

Abstract: In one embodiment, a feature map of an image having h×w pixels and a patch having one or more pixels of the image are received. The patch has been processed by a first set of layers of a convolutional neural network and contains an object centered within the patch. The patch is then processed using the feature map and one or more pixel classifiers of a classification layer of a deep-learning model, where the classification layer includes h×w pixel classifiers, with each pixel classifier corresponding to a respective pixel of the patch. Each of the pixel classifiers used to process the patch outputs a respective value indicating whether the corresponding pixel belongs to the object centered in the patch.

Abstract: In one embodiment, a method includes a computing system accessing a training image. The system may generate a feature map for the training image using a first neural network. The system may identify a region of interest in the feature map and generate a regional feature map for the region of interest based on sampling locations defined by a sampling region. The sampling region and the region of interest may correspond to the same region in the feature map. The system may generate an instance segmentation mask associated with the region of interest by processing the regional feature map using a second neural network. The second neural network may be trained using the instance segmentation mask. Once trained, the second neural network is configured to generate instance segmentation masks for object instances depicted in images.

Abstract: Systems, methods, and non-transitory computer-readable media can generate an initial alpha mask for an image based on machine learning techniques. A plurality of uncertain pixels is defined in the initial alpha mask. For each uncertain pixel in the plurality of uncertain pixels, a binary value is assigned based on a nearest certain neighbor determination.

Abstract: A method of object detection includes receiving a first image taken from a first perspective by a first camera and receiving a second image taken from a second perspective, different from the first perspective, by a second camera. Each pixel in the first image is offset relative to a corresponding pixel in the second image by a predetermined offset distance resulting in offset first and second images. A particular pixel of the offset first image depicts a same object locus as a corresponding pixel in the offset second image only if the object locus is at an expected object-detection distance from the first and second cameras. The method includes recognizing that a target object is imaged by the particular pixel of the offset first image and the corresponding pixel of the offset second image.

Abstract: In one embodiment, a feature map of an image having h×w pixels and a patch having one or more pixels of the image are received. The patch has been processed by a first set of layers of a convolutional neural network and contains an object centered within the patch. The patch is then processed using the feature map and one or more pixel classifiers of a classification layer of a deep-learning model, where the classification layer includes h×w pixel classifiers, with each pixel classifier corresponding to a respective pixel of the patch. Each of the pixel classifiers used to process the patch outputs a respective value indicating whether the corresponding pixel belongs to the object centered in the patch.

Abstract: In one embodiment a plurality of patches of an image are processed, using a first set of layers of a convolutional neural network, to output a plurality of object proposals associated with the plurality of patches of the image. Each patch includes one or more pixels of the image. Each object proposal includes a prediction as to a location of an object in the respective patch. Using a second set of layers of the convolutional neural network, the plurality of object proposals outputted by the first set of layers are processed to generate a plurality of refined object proposals. Each refined object proposal includes pixel-level information for the respective patch of the image. The first layer in the second set of layers of the convolutional neural network takes as input the plurality of object proposals outputted by the first set of layers.

Abstract: In one embodiment, a plurality of patches of an image are processed using a first-pass of a first deep-learning model to generate object-level information for each of the patches. Each patch includes one or more pixels of the image. Using a second-pass of the first deep-learning model, a respective object proposal is generated for each of the plurality of patches of the image. The second-pass takes as input the first-pass output, and the generated respective object proposals comprise pixel-level information for each of the patches. Using a second deep-learning model, a respective score is computed for each object proposal. The second deep-learning model takes as input the first-pass output, and the object score includes a likelihood that the respective patch of the object proposal contains an entire object.

Abstract: In one embodiment, a plurality of patches of an image are processed using a first-pass of a first deep-learning model to generate object-level information for each of the patches. Each patch includes one or more pixels of the image. Using a second-pass of the first deep-learning model, a respective object proposal is generated for each of the plurality of patches of the image. The second-pass takes as input the first-pass output, and the generated respective object proposals comprise pixel-level information for each of the patches. Using a second deep-learning model, a respective score is computed for each object proposal. The second deep-learning model takes as input the first-pass output, and the object score includes a likelihood that the respective patch of the object proposal contains an entire object.

Abstract: In one embodiment, a plurality of patches of an image are processed using a first deep-learning model to detect a plurality of features associated with the first patch of the image. Each patch includes one or more pixels of the image. Using a second deep-learning model, a respective object proposal is generated for each of the plurality of patches of the image. The second deep-learning model takes as input the plurality of detected features associated with the respective patch of the image, and each object proposal includes a prediction as to a location of an object in the patch. Using a third deep-learning model, a respective score is computed for each object proposal generated using the second deep-learning model. The third deep-learning model takes as input the plurality of detected features associated with the respective patch of the image, and the object score may include a likelihood that the patch contains an entire object.

Abstract: In one embodiment a plurality of patches of an image are processed, using a first set of layers of a convolutional neural network, to output a plurality of object proposals associated with the plurality of patches of the image. Each patch includes one or more pixels of the image. Each object proposal includes a prediction as to a location of an object in the respective patch. Using a second set of layers of the convolutional neural network, the plurality of object proposals outputted by the first set of layers are processed to generate a plurality of refined object proposals. Each refined object proposal includes pixel-level information for the respective patch of the image. The first layer in the second set of layers of the convolutional neural network takes as input the plurality of object proposals outputted by the first set of layers.

Abstract: A method of object detection includes receiving a first image taken from a first perspective by a first camera and receiving a second image taken from a second perspective, different from the first perspective, by a second camera. Each pixel in the first image is offset relative to a corresponding pixel in the second image by a predetermined offset distance resulting in offset first and second images. A particular pixel of the offset first image depicts a same object locus as a corresponding pixel in the offset second image only if the object locus is at an expected object-detection distance from the first and second cameras. The method includes recognizing that a target object is imaged by the particular pixel of the offset first image and the corresponding pixel of the offset second image.

Abstract: Systems, methods, and non-transitory computer-readable media can generate an initial alpha mask for an image based on machine learning techniques. A plurality of uncertain pixels is defined in the initial alpha mask. For each uncertain pixel in the plurality of uncertain pixels, a binary value is assigned based on a nearest certain neighbor determination.

Abstract: A method of object detection includes receiving a first image taken by a first stereo camera, receiving a second image taken by a second stereo camera, and offsetting the first image relative to the second image by an offset distance selected such that each corresponding pixel of offset first and second images depict a same object locus if the object locus is at an assumed distance from the first and second stereo cameras. The method further includes locating a target object in the offset first and second images.

Abstract: Edges are detected in a digital image including a plurality of pixels. For each of the plurality of pixels, a plurality of different edge assessments are made for that pixel. Each different edge assessment considers that pixel in a different position of a different pixel patch. The different edge assessments for each pixel are aggregated.

Abstract: In one embodiment, a plurality of patches of an image are processed using a first deep-learning model to detect a plurality of features associated with the first patch of the image. Each patch includes one or more pixels of the image. Using a second deep-learning model, a respective object proposal is generated for each of the plurality of patches of the image. The second deep-learning model takes as input the plurality of detected features associated with the respective patch of the image, and each object proposal includes a prediction as to a location of an object in the patch. Using a third deep-learning model, a respective score is computed for each object proposal generated using the second deep-learning model. The third deep-learning model takes as input the plurality of detected features associated with the respective patch of the image, and the object score may include a likelihood that the patch contains an entire object.

Abstract: The techniques and systems described herein are directed to isolating part-centric motion in a visual scene and stabilizing (e.g., removing) motion in the visual scene that is associated with camera-centric motion and/or object-centric motion. By removing the motion that is associated with the camera-centric motion and/or the object-centric motion, the techniques are able to focus motion feature extraction mechanisms (e.g., temporal differencing) on the isolated part-centric motion. The extracted motion features may then be used to recognize and/or detect the particular type of object and/or estimate a pose or position of a particular type of object.

Abstract: In one embodiment, a method may include receiving a first content item. A first embedding of the first content item may be determined and may corresponds to a first point in an embedding space. The embedding space may include a plurality of second points corresponding to a plurality of second embeddings of second content items. The embeddings are determined using a deep-learning model. The points are located in one or more clusters in the embedding space, which are each associated with a class of content items. Locations of points within clusters may be based on one or more attributes of the respective corresponding content items. Second content items that are similar to the first content item may be identified based on the locations of the first point and the second points and on particular clusters that the second points corresponding to the identified second content items are located in.

Abstract: Edges are detected in a digital image including a plurality of pixels. For each of the plurality of pixels, a plurality of different edge assessments are made for that pixel. Each different edge assessment considers that pixel in a different position of a different pixel patch. The different edge assessments for each pixel are aggregated.

Abstract: Technologies pertaining to object detection are described herein. A cascaded classifier executes over subwindows of an image in a plurality of stages. A crosstalk cascade is employed to reject subwindows as being candidates for including an object that is desirably detected, where the crosstalk cascade is a combination of multiple cascades.