Trust Your Good Friends: Source-Free Domain Adaptation by Reciprocal Neighborhood Clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation (DA) addresses discrepancies between source and target data domains.
• Traditional DA methods often require access to source data, which is frequently unavailable due to privacy or IP concerns.
• Source-free domain adaptation (SFDA) is a challenging problem where adaptation occurs without source data.

Purpose of the Study:

• To develop an effective SFDA method that adapts a pre-trained source model to a target domain without access to source data.
• To exploit the intrinsic cluster structure within target data for improved domain adaptation.
• To enhance label consistency and model robustness by considering local data structures.

Main Methods:

• The proposed method captures the intrinsic structure of target data by defining local affinity.
• It encourages label consistency among target samples with high local affinity, considering reciprocal neighbors and expanded neighborhoods.
• Target sample density is incorporated to mitigate the impact of outliers.

Main Results:

• The inherent structure of target features is a crucial information source for domain adaptation.
• Local neighborhood information (local, reciprocal, and expanded) effectively captures this structure.
• State-of-the-art performance is achieved on 2D image and 3D point cloud recognition datasets.

Conclusions:

• The proposed SFDA method effectively utilizes the local structure of target data for adaptation.
• Leveraging target data's intrinsic properties, such as clustering and density, is key to successful SFDA.
• The method demonstrates strong generalization capabilities across different data types and domains.