.png){kind=icon}
Return
Exploiting historical knowledge for source-free object detection via class prototype alignment
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.134714.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Source-Free Object Detection (SFOD) transfers a pre-trained detector to the unlabeled target domain without access to any training data from the source domain. Most existing SFOD methods follow the mean-teacher frame work, in which the student and teacher models are updated alternately. Specifically, the framework first optimizes a student model using filtered predictions of the teacher model as pseudo labels and then updates the teacher model with the parameters of the student model. However, the alternative updates can lead to catastrophic train ing collapse due to the unconstrained positive feedback of incorrect supervisory signals. To tackle the training collapse problem, we propose Class Prototype Alignment (CPA), a novel approach that uses class prototypes to leverage historical knowledge and mitigate the impact of incorrect supervisory signals. The proposed CPA frame work is composed of the Proposal Filter (PF) and Class Prototype Learning (CPL). The PF component selects well-localized proposals to obtain high-quality instance representations for prototype construction, while CPL iteratively updates class prototypes using historical knowledge accumulated from both the teacher and student models. Based on the obtained class prototypes, we propose the prototype contrastive loss and knowledge distilla tion loss to enhance the training stability of the model. Extensive experiments across diverse datasets demonstrate that our proposed method achieves state-of-the-art or competitive performance across different SFOD settings.
Keywords:
Source-free domain adaptation
Transfer learning
Unsupervised learning
Object detection
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
Cited Papers
No cited papers available