Abstract: A method of detecting an action, an electronic device, and a storage medium. A method can include: performing a temporal action proposal on at least one target feature data obtained by a feature extraction on a plurality of target frame data of a target resource, so as to obtain at least one first candidate action proposal information; classifying target feature data corresponding to at least one first candidate action proposal interval included in the first candidate action proposal information, so as to obtain at least one classification confidence level corresponding to the at least one first candidate action proposal interval; and determining an action detection result for at least one action segment contained in the target resource according to the at least one classification confidence level corresponding to the at least one first candidate action proposal interval, wherein the action detection result includes an action category and an action period.

Abstract: The present disclosure discloses a method and apparatus for extracting a video clip, relates to the field of artificial intelligence technology such as video processing, audio processing, and cloud computing. The method includes: acquiring a video, and extracting an audio stream in the video; determining a confidence that audio data in each preset period in the audio stream comprises a preset feature; and extracting a target video clip corresponding to a location of a target audio clip in the video; wherein the target audio clip is an audio clip within a continuous preset period, and has a confidence that the audio data includes the preset feature, which is larger than a preset confidence threshold. This method may improve the accuracy of extracting a video clip.

Abstract: A method of detecting an action, an electronic device, and a storage medium. A method can include: performing a temporal action proposal on at least one target feature data obtained by a feature extraction on a plurality of target frame data of a target resource, so as to obtain at least one first candidate action proposal information; classifying target feature data corresponding to at least one first candidate action proposal interval included in the first candidate action proposal information, so as to obtain at least one classification confidence level corresponding to the at least one first candidate action proposal interval; and determining an action detection result for at least one action segment contained in the target resource according to the at least one classification confidence level corresponding to the at least one first candidate action proposal interval, wherein the action detection result includes an action category and an action period.

Abstract: In one example, a system for real-time three-dimensional (3D) calibration includes a processor detecting mark corners at two-dimensional (2D) coordinates in an image coordinate system of an image of a specified pattern. The processor determines 3D coordinates of the mark corners in a world coordinate system based on the pattern. The 3D coordinates correspond to the 2D coordinates. The processor determines baseline depth values corresponding to original depth values for proper points on a mark plane, based on the 2D coordinates and the 3D coordinates. The processor generates a fitting model based on the baseline depth values.

Abstract: The present disclosure provides a living-body detection method for a face. The method includes: when an object to be detected is not illuminated by a reference light source, performing image collection on an area to be detected of the object to be detected to obtain a first image; performing illumination on the object to be detected by utilizing the reference light source, and performing image collection on the area to be detected of the object to be detected to obtain a second image; and judging whether there is a bright spot reflected by a cornea of the object to be detected according to a difference between the first image and the second image to generate a judged result, and determining whether the object to be detected is a non-living body according to the judged result.

Abstract: The present disclosure discloses a method and apparatus for extracting a video clip, relates to the field of artificial intelligence technology such as video processing, audio processing, and cloud computing. The method includes: acquiring a video, and extracting an audio stream in the video; determining a confidence that audio data in each preset period in the audio stream comprises a preset feature; and extracting a target video clip corresponding to a location of a target audio clip in the video; wherein the target audio clip is an audio clip within a continuous preset period, and has a confidence that the audio data includes the preset feature, which is larger than a preset confidence threshold. This method may improve the accuracy of extracting a video clip.

Abstract: The present disclosure provides a living-body detection method for a face. The method includes: when an object to be detected is not illuminated by a reference light source, performing image collection on an area to be detected of the object to be detected to obtain a first image; performing illumination on the object to be detected by utilizing the reference light source, and performing image collection on the area to be detected of the object to be detected to obtain a second image; and judging whether there is a bright spot reflected by a cornea of the object to be detected according to a difference between the first image and the second image to generate a judged result, and determining whether the object to be detected is a non-living body according to the judged result.

Abstract: In one example, a system for real-time three-dimensional (3D) calibration includes a processor detecting mark corners at two-dimensional (2D) coordinates in an image coordinate system of an image of a specified pattern. The processor determines 3D coordinates of the mark corners in a world coordinate system based on the pattern. The 3D coordinates correspond to the 2D coordinates. The processor determines baseline depth values corresponding to original depth values for proper points on a mark plane, based on the 2D coordinates and the 3D coordinates. The processor generates a fitting model based on the baseline depth values.

Abstract: In some examples, a camera calibration method for calibrating a plurality of cameras includes determining a relationship between each camera and a unique coordinate system. The calibration method also includes determining a positional relationship between all of the plurality of cameras based on the relationships between each of the cameras and the unique coordinate system.