Abstract: The technology relates to methods and systems for performing two-stage suppression of bounding boxes generated during object detection techniques for digital images. The two-stage suppression includes a per-class suppression stage and a class-agnostic suppression stage. In an example method, preliminary bounding boxes are generated for multiple objects in a digital image. A first subset of bounding boxes is selected by performing a per-class suppression of the preliminary bounding boxes. A second subset of bounding boxes is selected by performing a class-agnostic suppression of the first subset of bounding boxes. Based on the second subset of bounding boxes, at least one of an enriched image or a video index is generated.

Abstract: Examples of the present disclosure describe systems and methods for track aware object detection. In examples, image content comprising one or more objects is received. Frames in the image content are identified. Candidate bounding boxes are created around objects to be tracked in the frames and a confidence score is assigned to each candidate bounding box. The candidate bounding boxes for each object are compared to a predicted bounding box that is generated based on a current track for the object. Candidate bounding boxes that are determined to be similar to the predicted bounding box and/or that exceed a confidence score threshold are selected. The selected candidate bounding boxes are filtered until a single candidate bounding box that is most representative of each object to be tracked remains. The frame comprising the representative bounding box for each object is then added to a current track for the object.

Abstract: An object tracking tool integrates scene transition detection and/or dynamic queue resizing. By integrating shot transition detection, the object tracking tool can change which operations are performed depending on whether a shot transition has been detected. For example, if a shot transition is not detected, lower-complexity interpolation operations can be performed to determine spatial information for objects, instead of using higher-complexity object detection operations, which can reduce computational complexity. As another example, depending on whether a shot transition has been detected, the object tracking tool can adjust operations performed when associating identifiers with objects, which can improve accuracy of object tracking operations. With dynamic queue resizing, an object tracking tool can selectively adjust the maximum size of a queue used to store frames for object tracking.

Abstract: There is provided a method, comprising feeding a medical image into a detector component trained on a first training dataset of medical images annotated with ground truth boxes depicting a visual finding, obtaining boxes each associated with a respective box score indicative of likelihood of the visual finding, converting each respective box into a respective patch, feeding patches into a patch classifier trained on a second training dataset that includes patches extracted from the ground truth box labels of the first training dataset, wherein a patch score for a patch corresponds to a box score obtained from a box corresponding to the patch, obtaining patch scores indicative of likelihood of the visual finding being depicted, and computing a dot product of the box scores and the patch scores, and providing the dot product as an image-level indication of likelihood of the visual finding being depicted in the medical image.

Abstract: There is provided a method, comprising feeding a medical image into a detector component trained on a first training dataset of medical images annotated with ground truth boxes depicting a visual finding, obtaining boxes each associated with a respective box score indicative of likelihood of the visual finding, converting each respective box into a respective patch, feeding patches into a patch classifier trained on a second training dataset that includes patches extracted from the ground truth box labels of the first training dataset, wherein a patch score for a patch corresponds to a box score obtained from a box corresponding to the patch, obtaining patch scores indicative of likelihood of the visual finding being depicted, and computing a dot product of the box scores and the patch scores, and providing the dot product as an image-level indication of likelihood of the visual finding being depicted in the medical image.