Foreground Confusion under Domain Shift: The Hidden Bottleneck in Source‑Free Domain Adaptive Object Detection

ICLR 2026 Conference Submission12977 Authors

Keywords: Source-Free Domain Adaptation, Domain shift, Object Detection

Abstract: Source-Free Domain Adaptive Object Detection (SFOD) adapts detectors to new domains without source data, which is vital when privacy or storage constraints apply. SFOD is hindered by two key challenges: unreliable pseudo-labels, and foreground-background confusion, which occurs when domain shift induces spurious background activations that degrade localization and, in turn, classification. We introduce FOCUS-SFOD, a lightweight, architecture-agnostic framework with two complementary losses: CLEAN (Consistency Loss for Eliminating Activation Noise) mitigates foreground-background confusion by aligning channel-mean maps with simple foreground priors, improving localization; PAERL (Peak-Adjusted Entropy-Regularized Loss) reduces sensitivity to noisy pseudo class-labels by down-weighting trivial teacher-student agreements, encouraging learning on harder or underrepresented categories. To the best of our knowledge, we are the first to formalize foreground-background confusion in SFOD and provide a risk-bound analysis linking CLEAN and PAERL to tighter localization and classification errors. Across strong baselines and diverse shifts, FOCUS-SFOD delivers consistent gains of up to \(+3.9\) mAP, with zero inference overhead.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 12977