Data processing system, object detection method, and apparatus thereof
This application discloses a data processing system, an object detection method, and an apparatus thereof, and is applied to the artificial intelligence field. In this application, a second feature map generation unit introduces shallow texture detail information of an original feature map (a plurality of first feature maps generated by a convolutional processing unit) into a deep feature map (a plurality of second feature maps generated by a first feature map generation unit), to generate a plurality of third feature maps, and the third feature map including the shallow rich texture detail information is used as entered data on which a detection unit is to perform target detection.
1 . A cognitive network training method, the method comprising:
obtaining a pre-annotated detection box of a target object in an image;
obtaining a target detection box corresponding to the image and a first cognitive network, the target detection box identifying the target object, the target detection box comprising a first corner point and a first center point, the pre-annotated detection box comprising a second corner point and a second center point, the first corner point and the second corner point are two endpoints of a diagonal of a rectangle, and the position difference is further positively related to a center point position difference between the first center point and the second center point in the image and is negatively related to a length between the first corner point and the second corner point; and
performing iterative training on the first cognitive network based on a loss function and outputting a second cognitive network, the loss function relating to an intersection-over-union (IoU) between the pre-annotated detection box and the target detection box.
2 . The method according to claim 1 , wherein the loss function is further related to a shape difference between the target detection box and the pre-annotated detection box, and the shape difference is negatively related to an area of the pre-annotated detection box.
3 . The method according to claim 1 , wherein the loss function is further related to a position difference between the target detection box and the pre-annotated detection box in the image, and the position difference is negatively related to the area of the pre-annotated detection box; or the position difference is negatively related to an area of a minimum bounding rectangle of a convex hull of the pre-annotated detection box and the target detection box.
4 . The method according to claim 1 , wherein the loss function comprises a target loss term related to the position difference, and the target loss term changes with the position difference; and
a change rate of the target loss term is greater than a first preset change rate when the position difference is greater than a preset value; and/or the change rate of the target loss term is less than a second preset change rate when the position difference is less than the preset value.