Abstract
One-Shot Object Detection (OSOD) aims to identify objects of novel classes in a target image given only a single query instance. However, current OSOD frameworks often suffer from base-class bias, where the model erroneously prioritizes familiar categories seen during training, and generalization degradation due to over-parameterized correlation modules. This paper introduces the Base-class Suppression with Prior Guidance (BSPG) network to address these challenges. By incorporating a Base-class Suppression (BcS) module to filter out familiar distractors and a non-parametric Prior Guidance (PG) module to provide unbiased search hints, BSPG achieves state-of-the-art performance on MS COCO and PASCAL VOC benchmarks.
Introduction
One-Shot Object Detection (OSOD) represents a significant challenge in computer vision: the ability to detect objects from previously unseen "novel" classes using just one example. While traditional object detectors excel at identifying a fixed set of "base" classes, OSOD requires a model to generalize its understanding of objectness and similarity to new domains.
Existing OSOD methods typically utilize a Siamese architecture to correlate features between a query image and a target image. However, two major issues persist. First, Base-Class Bias causes models to favor familiar categories. For instance, when searching for a novel animal in a scene containing a person, a model trained on "person" as a base class may mistakenly detect the person as the target. Second, Generalization Degradation occurs when complex, learnable matching modules overfit to the specific appearances of base classes, failing to adapt to the diverse characteristics of novel objects.
Methodology
The proposed Base-class Suppression with Prior Guidance (BSPG) network addresses these issues by explicitly modeling what to ignore and how to search without bias.
Base-class Suppression (BcS) Module
The BcS module serves as a "familiarity filter." It leverages a pre-trained Base-class Predictor to identify objects belonging to known categories within the target image. When the novel-class predictor generates proposals, the BcS module checks if these proposals overlap significantly with high-confidence base-class detections. If a proposal is likely a base-class object, its score is suppressed. This mechanism effectively removes "distractors" that would otherwise cause false positives due to base-class bias.
Prior Guidance (PG) Module
To improve localization without introducing learnable bias, the PG module generates a non-parametric Prior Map. By calculating the cosine similarity between high-level features of the query and target images, the model creates a heatmap of potential matches. This map is used to re-weight features in the Region of Interest (RoI) head, guiding the detector toward regions that structurally resemble the query, regardless of their specific class.
Training Strategy
The network is trained in two stages. First, the base-class predictor is trained on standard detection tasks. Second, these weights are frozen while the novel-class predictor and suppression parameters are optimized using an episodic training strategy, ensuring the model learns to effectively utilize the suppression and guidance signals.
Experiments and Results
The BSPG network was evaluated on the MS COCO and PASCAL VOC datasets, following standard OSOD protocols.
Quantitative Performance
On the MS COCO dataset, BSPG achieved an average of 27.3% AP50 across novel classes, representing a 1.9% improvement over the previous state-of-the-art, BHRL. In Few-Shot Object Detection (FSOD) scenarios, the model demonstrated even greater gains, outperforming competitors like DAnA by up to 5.0% in 5-shot settings. These results highlight the robustness of the suppression and guidance mechanisms across varying amounts of support data.
Qualitative Analysis
Visual comparisons demonstrate the effectiveness of the BcS module. While baseline models often trigger on "familiar" objects like people or cars when tasked with finding novel animals, BSPG successfully suppresses these detections.
Discussion and Conclusion
The success of BSPG suggests a paradigm shift in OSOD: rather than solely focusing on improving matching algorithms, researchers should also consider how to mitigate the inherent biases of the training process. By "subtracting" known information through the BcS module and providing "unbiased" hints via the PG module, BSPG achieves superior generalization.
However, the approach is not without limitations. The effectiveness of the BcS module is sensitive to the confidence threshold ($\alpha$), which may require tuning for different datasets. Additionally, the dual-predictor architecture increases computational complexity during inference. Future work may explore more integrated ways to utilize base-class knowledge without requiring a separate, full-scale predictor branch.
In conclusion, BSPG provides a powerful framework for bias-free one-shot detection, setting a new benchmark for the field and opening new avenues for addressing model bias in open-vocabulary tasks.
Summary updated