A target marking method, apparatus, computer equipment, and computer storage medium
By using a cascaded approach of YOLO and Sa2VA models, combined with polygon vertex simplification and incremental training, the recall and boundary consideration issues in automatic labeling of small targets are resolved. This achieves high-quality, reproducible, and scalable label generation, reducing manual costs and errors.
CN122134997APending Publication Date: 2026-06-02FIBOCOM WIRELESS
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FIBOCOM WIRELESS
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134997A_ABST
Abstract
This application discloses a target labeling method, apparatus, computer device, and computer storage medium, relating to the field of image processing technology. The method includes: reasoning about an image using a first model to obtain an instance mask; determining the circumscribed polygon of the instance mask; calling a second model to perform cue-based segmentation on the circumscribed polygon to obtain candidate masks; calculating the intersection-union ratio (IU) of the instance mask and the candidate masks, and calculating the union of candidate masks with IU greater than or equal to a preset threshold; and outputting a label based on the union. This method can balance recall and boundary conditions, and can achieve high-quality, reproducible, and scalable automatic labeling of small targets with low manual costs.
Need to check novelty before this filing date? Find Prior Art