Source-Free Object Detection With Detection Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Source-Free Object Detection (SFOD) facilitates knowledge transfer to unsupervised domains without source data.
• Existing SFOD methods often lack tailored adaptations for advanced architectures like Detection Transformers (DETR).

Purpose of the Study:

• To introduce FRANCK, a novel SFOD framework specifically designed for query-centric feature enhancement in DETR models.
• To improve the robustness and generalization of object detection models in unsupervised target domains.

Main Methods:

• FRANCK integrates four key components: Objectness Score-based Sample Reweighting (OSSR), Contrastive Learning with Matching-based Memory Bank (CMMB), Uncertainty-weighted Query-fused Feature Distillation (UQFD), and a Dynamic Teacher Updating Interval (DTUI) self-training pipeline.
• OSSR reweights detection loss using attention-based objectness scores.
• CMMB enhances class-wise contrastive learning via multi-level feature memory banks.
• UQFD improves feature distillation through query feature fusion and prediction quality reweighting.

Main Results:

• FRANCK effectively adapts source-pre-trained DETR models to target domains.
• The proposed method demonstrates enhanced robustness and generalization capabilities.
• Experiments on multiple benchmarks show state-of-the-art performance, validating FRANCK's effectiveness.

Conclusions:

• FRANCK offers a specialized and effective solution for Source-Free Object Detection within the DETR architecture.
• The framework significantly advances the capabilities of unsupervised domain adaptation for object detection.
• FRANCK's components collectively contribute to superior performance and compatibility with DETR-based SFOD models.