META-LEARNING STRATEGIES FOR FEW-SHOT OBJECT DETECTION: A COMPARATIVE ANALYSIS OF OPTIMIZATION-BASED AND METRIC-BASED APPROACHES

Few-shot object detection (FSOD) requires a model to localize and classify novel object categories from only a handful of annotated examples, while retaining performance on previously learned base categories. This paper examines meta-learning as a mechanism for achieving this goal, contrasting optimization-based strategies (e.g., MAML-style parameter adaptation) with metric-based strategies (e.g., prototypical embedding matching). We further evaluate a hybrid configuration in which a meta-learned backbone is paired with a dedicated second-stage classifier that refines coarse region proposals into fine-grained category labels — directly extending the adaptive few-shot framework developed in our related work. Using a standard MS-COCO / PASCAL VOC few-shot evaluation protocol, we compare the four configurations across K = 1, 5, 10, and 30 shots. Results indicate that optimization-based meta-learning generalizes faster to unseen categories than a naive fine-tuning baseline, and that the addition of a cosine-similarity second-stage classifier consistently improves novel-class average precision (nAP) across all shot regimes. We discuss the trade-offs between adaptation speed, annotation cost, and computational overhead, and outline directions for combining meta-learning with dynamic, second-stage fine-grained categorization.