Beyond Primitive Reuse: Primitive Discovery and Adaptive Modulation for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) is essential for adapting to open-world environments. Humans can rapidly learn novel concepts by decomposing objects into key components and recombining familiar components through associative memory. Inspired by this ability, compositional approaches reuse base-class primitives or update the decomposition module in subsequent sessions. However, whole-primitive reuse may overlook discriminative details, while jointly learning decomposition and classification is challenging with few examples. To address this limitation, we propose Primitive Discovery and Adaptive Modulation (PDAM) for FSCIL. Specifically, Object-centric Primitive Discovery (OPD) module learns to decompose objects into primitives from base-class images and transfers this capability to novel categories. However, the decomposition alone does not ensure effective recognition, as the resulting primitives contain both shared appearance and potentially discriminative details whose relevance is difficult to establish from few examples. We therefore introduce a Primitive-guided Adaptive Modulation (PAM) module that adaptively modulates features across regions and channels to support discrimination. Experiments demonstrate that our approach achieves state-of-the-art performance on three fine-grained FSCIL benchmarks.