US 12,327,388 B2
Multi-camera domain adaptive object detection system and detection method thereof
Peggy Joy Lu, Taichung (TW); and Jen-Hui Chuang, Hsinchu (TW)
Assigned to National Yang Ming Chiao Tung University, Hsinchu (TW)
Filed by National Yang Ming Chiao Tung University, Hsinchu (TW)
Filed on Oct. 7, 2022, as Appl. No. 17/962,388.
Claims priority of application No. 111126368 (TW), filed on Jul. 13, 2022.
Prior Publication US 2024/0020941 A1, Jan. 18, 2024
Int. Cl. G06V 10/25 (2022.01); G06V 10/774 (2022.01); G06V 20/52 (2022.01)
CPC G06V 10/25 (2022.01) [G06V 10/774 (2022.01); G06V 20/52 (2022.01); G06V 2201/07 (2022.01)] 17 Claims
OG exemplary drawing
 
1. A multi-camera domain adaptive object detection system, which is suitable for training a global model, comprising:
a server-end device including a target camera for obtaining one target data, wherein the target data includes an image data captured by the target camera; and
a plurality of client-end devices electrically connecting the server-end device, wherein each of the client-end devices comprises a source camera for obtaining one corresponding source data;
wherein, the object detection system executes at least following processing procedures:
processing procedure one: the server-end device transmits the global model to each of the client-end devices;
processing procedure two: each of the client-end devices trains the received global model according to the image data captured by the target camera and the corresponding source data so as to obtain a trained model;
processing procedure three: each of the client-end devices inputs the image data captured by the target camera into the trained model, extracts feature values to obtain one corresponding feature data, and transmits each of the corresponding feature data to the server-end device;
processing procedure four: the server-end device assembles the received feature data from the client-end devices to obtain a plurality of loss functions; and
processing procedure five: the server-end device trains the global model according to the target data by the loss functions to obtain a trained global model,
wherein, before starting the processing procedure one, the server-end device transmits the image data captured by the target camera to each of the client-end devices.