GiPL: Generative augmented iterative Pseudo-Labeling for Cross-Domain Few-Shot Object Detection

Researchers propose GiPL, a two-branch machine learning framework that combines iterative pseudo-labeling with generative data augmentation to improve cross-domain few-shot object detection using vision-language models. The method demonstrates significant performance improvements on three benchmark datasets, addressing critical challenges in fine-tuning with limited target-domain samples.

GiPL tackles a fundamental problem in computer vision: enabling object detection systems to work effectively across different domains with minimal labeled examples. The framework addresses two interconnected challenges that plague few-shot learning systems. First, training data scarcity creates insufficient signal for effective model optimization. Second, the extreme limitation of target-domain samples typically leads to severe overfitting, where models memorize rather than generalize.

The solution employs a clever two-pronged approach. The first branch leverages zero-shot inference capabilities of foundation models to generate pseudo-labels on support sets, creating synthetic annotations that expand available training signals without additional human labeling effort. By iteratively refining predictions, the system progressively improves label quality while maximizing use of existing data. The second branch uses generative models to synthesize realistic, domain-aligned training images with multiple objects, artificially expanding the dataset while maintaining domain relevance.

This advancement holds practical implications for computer vision applications in autonomous vehicles, robotics, and surveillance systems where real-world deployment domains often differ significantly from training data. The consistent performance gains across three challenging datasets—RUOD, CARPK, and CarDD—across varying shot settings (1, 5, and 10-shot scenarios) demonstrate robust generalization.

The research exemplifies how foundation models can enable more efficient learning with limited supervision. As organizations increasingly seek to deploy vision systems in new domains without extensive relabeling efforts, such methods become operationally valuable. Future work likely explores scaling these approaches to more complex scenarios and investigating how pseudo-label quality affects long-term model robustness.

• →GiPL combines iterative pseudo-labeling and generative augmentation to overcome few-shot object detection limitations
• →The framework demonstrates significant performance gains across three challenging datasets under extreme data scarcity conditions
• →Vision-language foundation models enable zero-shot inference to generate synthetic training signals without manual annotation
• →Generative data augmentation synthesizes domain-aligned images to suppress overfitting in limited-sample scenarios
• →The approach has practical applications for deploying object detection systems across different domains with minimal labeled data