Open vocabulary man-machine interaction detection method based on calibration diffusion model

By adopting a two-stage training framework based on a calibration diffusion model, the problems of noise and global feature bias in open-vocabulary human-computer interaction detection are solved, the detection accuracy and generalization ability are improved, and the efficient recognition of novel interactions is achieved.

CN120949945AActive Publication Date: 2025-11-14UNIV OF ELECTRONICS SCI & TECH OF CHINA
4 Cites 0 Cited by

Patent Information

Application Number
CN202511471311.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-15
Publication Date
2025-11-14
Estimated Expiration
2045-10-15

AI Technical Summary

Technical Problem

Existing technologies have poor performance in detecting unseen interaction categories in open-vocabulary human-computer interaction due to inherent noise in visual language models, global feature bias, and overfitting of spatial relationships.

Method used

A two-stage training framework based on a calibration diffusion model is adopted, including calibration and training of the teacher model and training of the HOI detector of the student model. Feature calibration is performed through a visual and text dual-branch structure, and end-to-end supervised optimization is performed by combining graph network learning and hybrid loss function.

Benefits of technology

It significantly improves the detection accuracy of open vocabulary, enhances the model's generalization ability and robustness, and enables a better understanding of the essential semantics and spatial relationships of interactions, achieving efficient detection of novel interactive scenarios.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses an open vocabulary man-machine interaction detection method based on a calibration diffusion model, and relates to the technical field of computer vision and artificial intelligence. The method adopts a two-stage teacher-student learning framework: a teacher model carries out calibration by injecting man-machine interaction priori knowledge into a generative diffusion model so as to generate low-noise and fine-grained visual features for a man-object interaction area; the student model is an end-to-end detection framework, overall optimization is carried out under the supervision of mixed loss, the secondary optimization problem caused by multi-stage independent training is avoided, the detection process is more efficient, and the performance is better. By introducing a human-object pair learning device and a local perception decoder and combining knowledge distillation of a teacher model, the method can understand essential semantics and spatial relationships of interaction more deeply, and has stronger generalization ability for novel and unseen interaction scenes. The method can effectively inhibit the internal noise of the pre-training model, and enables the features to be highly aligned with the HOI task.
Need to check novelty before this filing date? Find Prior Art