Bidirectional pairing architecture for object detection in video
AssigneeIntel Corporation
InventorsYan Hao, Zhi Yong Zhu, Lu Li, Ciyong Chen, Kun Yu
Life of this patent
Estimating the term…
Abstract
Techniques related to training and implementing a bidirectional pairing architecture for object detection are discussed. Such techniques include generating a first enhanced feature map for each frame of a video sequence by processing the frames in a first direction, generating a second enhanced feature map for frame by processing the frames in a second direction opposite the first, and determining object detection information for each frame using the first and second enhanced feature map for the frame.
Claims21
A system for performing object detection, comprising:
- a memory to store at least a portion of one or more of a current, a previous, and a subsequent frame of video;
- and
- a processor coupled to the memory, the processor to:
- generate a current frame initial feature map comprising object detection scoring for the current frame;
- detect, based on patch similarity, first and second paired patches between the current and previous frames, respectively, and third and fourth paired patches between the current and subsequent frames, respectively;
- increase a first prediction result in the current frame initial feature map for the first paired patch to a score of the second paired patch or increase a second prediction result in the current frame initial feature map for the third paired patch to a score of the fourth paired patch to generate an enhanced feature map;
- and
- determine one or more of an object detection localization, class, and confidence scoring for the current frame using the enhanced feature map.
A computer-implemented method for performing object detection, comprising:
- receiving a current, a previous, and a subsequent frame of video;
- generating a current frame initial feature map comprising object detection scoring for the current frame;
- detecting, based on patch similarity, first and second paired patches between the current and previous frames, respectively, and third and fourth paired patches between the current and subsequent frames, respectively;
- increasing a first prediction result in the current frame initial feature map for the first paired patch to a score of the second paired patch or increasing a second prediction result in the current frame initial feature map for the third paired patch to a score of the fourth paired patch to generate an enhanced feature map;
- and
- determining one or more of an object detection localization, class, and confidence scoring for the current frame using the enhanced feature map.
At least one non-transitory machine readable medium comprising a plurality of instructions that, in response to being executed on a computing device, cause the computing device to perform object detection by:
- receiving a current, a previous, and a subsequent frame of video;
- generating a current frame initial feature map comprising object detection scoring for the current frame;
- detecting, based on patch similarity, first and second paired patches between the current and previous frames, respectively, and third and fourth paired patches between the current and subsequent frames, respectively;
- increasing a first prediction result in the current frame initial feature map for the first paired patch to a score of the second paired patch or increasing a second prediction result in the current frame initial feature map for the third paired patch to a score of the fourth paired to generate an enhanced feature map;
- and
- determining one or more of an object detection localization, class, and confidence scoring for the current frame using the enhanced feature map.
Dependencies are inferred from the claim text.