Abstract
Few-Shot Cross-Domain (FSCD) object detection is a critical task for deploying robust computer vision systems in diverse environments where target domain data is scarce. Traditional Unsupervised Domain Adaptation (UDA) methods require large unlabeled datasets, which are often unavailable for specific weather conditions or geographic locations. We propose OA-FSUI2IT (Object-Aware Few-Shot Unsupervised Image-to-Image Translation), a framework that adapts detectors using as few as 10 unlabeled target images. By integrating discriminator augmentation, patch pyramid contrastive learning, and self-supervised consistency, OA-FSUI2IT generates high-quality synthetic training data that preserves source content while adopting target styles. Our method achieves state-of-the-art performance on major benchmarks, including transitions from clear to foggy weather and day to night.
Introduction
Object detection models trained on clear, daytime datasets often experience significant performance degradation when deployed in adverse conditions like heavy fog or nighttime. This "domain shift" is a major hurdle for autonomous systems. While Unsupervised Domain Adaptation (UDA) offers a solution by bridging the gap between source and target domains, it typically requires thousands of unlabeled target images.
In many practical scenarios, we only have access to a handful of images from the target domain. This leads to the Few-Shot Cross-Domain (FSCD) challenge. Existing Image-to-Image (I2I) translation techniques, such as CycleGAN or CUT, often fail in this low-data regime, producing artifacts or overfitting to the small target sample.
Methodology: OA-FSUI2IT
The OA-FSUI2IT framework addresses the limitations of few-shot translation by enhancing the Contrastive Unpaired Translation (CUT) baseline with three key components:
1. Discriminator Augmentation (DA)
To prevent the GAN's discriminator from overfitting to the limited (~10) target images, we employ "revertible" augmentations (e.g., flips, rotations, color transforms). This forces the model to learn general stylistic distributions rather than memorizing specific pixel patterns of the few available target samples.
2. Patch Pyramid Contrastive Learning (PPCL)
Standard contrastive methods often struggle with multi-scale objects. PPCL utilizes a pyramid of patches, analogous to the anchors in object detectors. This ensures that structural integrity is maintained for both large foreground objects and small, distant ones during the translation process.
3. Self-Supervised Content-Consistency (SSCC) Loss
To ensure that the translated images remain perfectly aligned with the original source labels, we introduce a consistency check. By applying transformations to the input and verifying that the translation remains structurally consistent after inverse transformation, we minimize geometric distortions that would otherwise invalidate bounding box annotations.
Experimental Results
We evaluated OA-FSUI2IT on three primary domain-shift benchmarks in a 10-shot setting:
- Normal-to-Foggy: Cityscapes $\rightarrow$ FoggyCityscapes.
- Day-to-Night: BDD100k Daytime $\rightarrow$ BDD100k Nighttime.
- Cross-Scene: Kitti $\rightarrow$ Cityscapes.
Our framework achieved significant improvements over baselines. In the Normal-to-Foggy task, OA-FSUI2IT reached 42.3 mAP, compared to 30.0 mAP for the source-only model and 39.5 mAP for the standard CUT baseline. In the Day-to-Night transition, the model achieved 30.5 mAP, a notable increase from the 26.4 mAP source-only baseline. Qualitative analysis confirms that our method produces cleaner, more realistic translations without the "checkerboard" artifacts common in other GAN-based approaches.
Conclusion and Future Work
OA-FSUI2IT demonstrates that high-quality image translation is possible even with extremely limited target data. By "re-skinning" labeled source data into the target style, we create a robust synthetic dataset for training detectors.
While the model significantly reduces missed detections, error analysis indicates a slight increase in false positives, likely due to the detector mistaking generated textures for objects. Future research will focus on refining texture generation to minimize these errors and exploring "one-shot" adaptation scenarios.
Summary updated