US 12,437,521 B2
Few-shot object detection method
Zhonghong Ou, Beijing (CN); Junwei Yang, Beijing (CN); Xiaoyang Kang, Beijing (CN); Jiawei Fan, Beijing (CN); Xie Yu, Beijing (CN); and Meina Song, Beijing (CN)
Assigned to BEIJING UNIVERSITY OF POSTS & TELECOMMUNICATIONS, Beijing (CN)
Appl. No. 18/551,919
Filed by BEIJING UNIVERSITY OF POSTS & TELECOMMUNICATIONS, Beijing (CN)
PCT Filed Sep. 8, 2022, PCT No. PCT/CN2022/117896
§ 371(c)(1), (2) Date Sep. 22, 2023,
PCT Pub. No. WO2023/109208, PCT Pub. Date Jun. 22, 2023.
Claims priority of application No. 202111535847.6 (CN), filed on Dec. 15, 2021.
Prior Publication US 2024/0177462 A1, May 30, 2024
Int. Cl. G06V 10/774 (2022.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/77 (2022.01); G06V 10/82 (2022.01)
CPC G06V 10/7753 (2022.01) [G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/7715 (2022.01); G06V 10/82 (2022.01)] 18 Claims
OG exemplary drawing
 
1. A few-shot object detection method, comprising:
sending a weight of a backbone network and a weight of a feature pyramid to a detection network, wherein the weight of the backbone network and the weight of the feature pyramid are derived from a visual representation backbone network generated by self-supervised training;
generating candidate regions, wherein the candidate regions are derived from a result of foreground-and-background view classification and regression of output features of the visual representation backbone network by a region proposal network;
generating candidate region features of a uniform size using a pooling operator based on the candidate regions, and performing location regression, content classification and fine-grained feature mining on the candidate region features of the uniform size;
establishing fine-grained positive sample pairs and negative sample pairs through the fine-grained feature mining, and performing comparative learning between fine-grained features of the candidate regions, wherein the fine-grained feature mining comprises a strategy of: performing region division on the uniform-sized candidate region features using an even division manner, extracting features of different regions after the division, assigning same labels to region division results from a same candidate region and assigning different labels to region division results from different candidate regions; and
generating a loss function according to the strategy in the fine-grained feature mining, and updating detection network parameters by calculating based on the loss function.