Synthetic‐to‐Real Few‐Shot Crowd Counting via Scene Adaptive Meta Learning

ABSTRACT

Crowd counting is a fundamental task in computer vision. However, acquiring annotated real‐world data is labour‐intensive and costly, making it challenging to train effective models. To tackle this, we introduce a new task termed synthetic‐to‐real few‐shot crowd counting (S2R‐FSCC), which leverages large‐scale synthetic data for training and only a small amount of real data for fine‐tuning. This setup requires models to rapidly adapt from the synthetic to the real domain under limited target supervision. To this end, we propose a scene adaptive meta‐learner (SAML) embedded within a meta‐learning framework. During the meta‐training phase, SAML learns from diverse synthetic tasks and dynamically adjusts its parameters according to the input scene, facilitating rapid cross‐domain adaptation. During the meta‐testing phase, the model is further fine‐tuned with limited real‐world samples using a teacher–student distillation strategy, whereby the teacher preserves transferable knowledge to guide the student's update and alleviate catastrophic forgetting. Extensive experiments on three real‐world datasets demonstrate that our meta‐training and meta‐testing strategies significantly enhance the model's adaptability under both 1‐shot and 5‐shot settings. These results highlight the effectiveness of synthetic‐to‐real few‐shot adaptation in reducing reliance on large‐scale annotated real‐world crowd data.