Abstract
Zero-Shot Object Detection (ZSD) aims to detect objects from classes not seen during training. Traditional inductive approaches often suffer from domain shift and model bias toward seen categories. This paper proposes the first transductive learning approach for ZSD, leveraging unlabeled target domain images during training. By employing a novel hybrid pseudo-labeling strategy—combining fixed labels for seen classes and dynamic labels for unseen classes—the method achieves a significant 77% relative improvement in recall on the MSCOCO dataset.
Introduction
Standard object detection models are confined to the specific categories present in their training data. Zero-Shot Object Detection (ZSD) seeks to overcome this by enabling models to locate and identify "unseen" objects (e.g., detecting a "bear" after training only on "dogs" and "cats").
ZSD faces two primary challenges: Domain Shift, where visual features of seen classes differ from unseen ones, and Model Bias, where the detector tends to misclassify unseen objects as similar-looking seen categories. While previous research focused on the inductive setting (training only on labeled data), this work advocates for a transductive setting. In this paradigm, the model observes unlabeled images from the target domain during training to learn the underlying distribution of the unseen world.
Proposed Methodology
The authors build their solution upon the RetinaNet architecture, introducing a Self-Learning Mechanism powered by Hybrid Pseudo-Labeling. This strategy bridges the gap between seen and unseen domains through two distinct phases:
- Fixed Pseudo-Labeling: An initial inductive model is trained on labeled data and used to predict labels for unlabeled target images. Predictions for "seen" classes are kept fixed to serve as stable anchors, preventing the model from forgetting established knowledge.
- Dynamic Pseudo-Labeling: As training progresses, the model's confidence in "unseen" classes increases. The system dynamically updates labels for these objects, allowing the decision boundaries to evolve.
- Unseen Score Maximization: To counteract bias, a specific loss term is introduced to maximize prediction scores for unseen classes, ensuring they are not erroneously categorized as background.
Key Insights and Analysis
The core advantage of transduction in ZSD is its ability to provide the "missing link" between semantic class names and visual textures. While we know the names of unseen classes, transductive learning allows the model to observe their physical shapes and textures in the test domain without explicit labels.
As shown in the analysis of projection scores, the transductive approach successfully increases the gap between similar and dissimilar unseen classes. This makes the model significantly more decisive when encountering novel objects, effectively breaking the inherent bias toward seen categories.
Experimental Results
The approach was evaluated on the MSCOCO (2014) and PASCAL VOC benchmarks. On MSCOCO, the model achieved a 34.9% relative improvement in mAP and a 77.1% relative improvement in Recall over state-of-the-art inductive models. In the Generalized ZSD (GZSD) setting, where both seen and unseen objects co-occur, the model maintained high performance, successfully detecting difficult classes like "bear" and "hot dog" in complex scenes where inductive baselines failed.
Conclusion and Future Work
This work demonstrates that transductive learning is a powerful tool for overcoming domain shift in ZSD. However, the requirement for unlabeled target data during training may limit its use in strictly open-world or real-time streaming scenarios. Future research will focus on improving the distinction between background noise and unseen objects, which remains a significant challenge in the detection pipeline.
Summary updated