TL;DR
This paper tackles the instability and "forgetting" problem in few-shot object detection (FSOD) by introducing a simple yet effective "sample normalization" technique. By stabilizing the representation of support data, their method significantly improves detection accuracy and consistency across different few-shot scenarios, even for previously seen "base" classes.
What Problem Does This Paper Address?
Few-shot object detection (FSOD) is crucial for real-world applications where labeled data is scarce. The goal is to train models that can recognize and locate objects from new classes using only a handful of examples. While meta-learning approaches have shown promise, they often suffer from two major issues: instability (performance varies wildly depending on the specific few-shot samples chosen) and forgetting (the model's performance on previously learned "base" classes degrades significantly when trained on new "novel" classes). The authors identify two key gaps contributing to these problems: a distribution gap between training and test data, and a gap between source and target domains in meta-learning, both of which hinder effective knowledge transfer.
Methods
The core contribution is Sample Normalization (SN), a technique applied to the support branch of meta-learning detectors. SN standardizes the support features to reside on a unit hypersphere. This aims to:
- Mitigate data volume disparities: By aligning support features from different stages (base training vs. fine-tuning), it improves meta-knowledge transfer and preserves base class performance.
- Reduce inconsistencies in few-shot data: It harmonizes variations within randomly sampled few-shot data, leading to more stable network weights and consistent performance.
Additionally, the paper incorporates Z-score normalization (ZN), followed by a linear projection, to address the hubness problem. Hubness, common in high-dimensional feature spaces, causes some data points to be disproportionately close to many others, potentially degrading detection performance. ZN helps to normalize these features and mitigate this issue. The authors use Meta-DETR as their backbone and integrate SN and ZN into its architecture.
Datasets
The experiments were conducted on the Pascal VOC dataset, a standard benchmark for object detection and FSOD, containing 20 categories. The authors used the VOC 07+12 training data for meta-training and the VOC 07 test data for evaluation. They specifically set up experiments with different combinations of target and source domain categories and varied the number of few-shot samples from 1 to 10.
Key Insights and Ideas
- The "forgetting" problem is deeply tied to data distribution gaps: The authors convincingly argue that the instability and forgetting in meta-learning FSOD stem from mismatches in how data is represented and sampled across different stages and domains.
- Simple normalization can be powerful: Instead of complex architectural changes, a straightforward normalization technique (SN) proves highly effective in stabilizing meta-learning. This highlights the importance of feature representation in FSOD.
- Hubness is a real concern in FSOD: The inclusion of Z-score normalization specifically to combat hubness underscores its detrimental impact on detection performance in high-dimensional feature spaces.
- Balancing novel and base class performance is achievable: The proposed method not only improves performance on novel classes but also significantly reduces the degradation of performance on base classes, a common trade-off in FSOD.
Key Results
The proposed method, "Ours," consistently outperforms the baseline Meta-DETR across various few-shot settings (1 to 10 shots) and class splits on the Pascal VOC dataset.
- Accuracy Improvements: Achieved up to +4.4 [email protected] and +5.3 mAR in a single run, and +4.8 [email protected] and +5.1 mAR on average over 10 random experiments.
- Stability: Demonstrated significantly smaller variance in detection performance (mAP and mAR) compared to Meta-DETR, indicating improved robustness across different few-shot samples. Table 3 shows a substantial reduction in variance, particularly for mAR.
- Base Class Performance: The method effectively alleviates the drop in performance of base classes, retaining strong detection ability. For instance, in Table 4, their method achieves higher mAP for base classes compared to Meta-DETR.
- Dense Object Detection: The method also showed excellent performance in challenging scenarios with dense or occluded objects, accurately identifying individual targets.
Limitations and Future Work
- Extreme Sample Variations: The authors acknowledge that in cases of extreme differences in few-shot samples (due to different random seeds), the network might still tend towards one extreme, indicating room for further improvement in balancing performance across such diverse inputs.
- Two-Stage Training Complexity: The current two-stage training strategy (meta-training and meta-fine-tuning) is described as complex and time-consuming.
- Future Directions: The authors plan to explore more effective training methods and investigate sample data distribution to further bridge the gap between few-shot and large-sample data. They also aim to develop approaches that might not require fine-tuning for detecting new categories.
Summary updated