Scale-Consistent and Temporally Ensembled Unsupervised Domain Adaptation for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation for Object Detection (UDA-OD) addresses domain shifts but struggles with small objects and inaccurate pseudo-labels.
• Existing UDA-OD methods often over-rely on classification confidence, leading to poor bounding box localization.

Purpose of the Study:

• To develop a novel UDA-OD framework enhancing cross-domain robustness and detection performance, particularly for small objects.
• To improve the accuracy of pseudo-label selection and bounding box localization in UDA-OD.

Main Methods:

• Introduced Cross-Scale Prediction Consistency (CSPC) for robust detection across multiple resolutions.
• Integrated Intra-Class Feature Consistency (ICFC) using contrastive learning to align feature representations.
• Developed Temporal Ensemble Pseudo-Label Selection (TEPLS) combining temporal stability and classification confidence for high-quality pseudo-labels.

Main Results:

• Achieved state-of-the-art performance on challenging UDA-OD benchmarks (Cityscapes, Sim10k, Virtual Mine).
• Demonstrated significant improvements in small object detection accuracy.
• Showcased enhanced overall cross-domain robustness compared to existing methods.

Conclusions:

• The proposed UDA-OD framework effectively addresses limitations of current approaches.
• The combination of scale consistency and advanced pseudo-label selection significantly boosts detection performance.
• The method offers a robust solution for real-world object detection challenges with domain shifts.