Self-adaptive deployment of synthetic-trained object detectors via test-time adaptation

Abstract

Deploying synthetic-trained object detectors in real environments is challenging under single-image, batch=1 inference, where pseudo-label noise and unstable normalization can undermine online adaptation. We present a deployment-oriented test-time adaptation framework for synthetic-to-real object detection that requires neither source data nor target labels. The framework combines entropy-guided adaptation triggering, IoU-filtered pseudo-labeling with augmentation consistency and greedy matching, and architecture-aware normalization updates for both CNN- and Transformer-based detectors. Trained on LookSIM synthetic data and evaluated on 3,500 real defense images spanning six target classes, the method improves $$\textrm{AP}@[.5\!:\!.95]$$ from 51.7 to 56.5 for RF-DETR and from 49.7 to 54.9 for YOLOv12-L while preserving real-time deployment on RTX 3090 and Jetson AGX Orin. Ablation results show that IoU filtering reduces pseudo-label false positives from 40% to 15%, memory-augmented normalization limits batch=1 degradation, and entropy-guided selection provides additional gains. Additional robustness analyses under severe occlusion and illumination stress show that source-side synthetic stress augmentation improves pseudo-label stability and reduces the loss of adaptation benefit in harder night and low-visibility regimes. Results on a held-out target condition and on the public RarePlanes benchmark further indicate that the proposed framework remains effective beyond both the dominant deployment subset and the specific LookSIM-to-defense setting. These findings strengthen the case for test-time adaptation as a deployment-oriented correction mechanism for synthetic-trained detectors when labeled real data remain limited.