Core Concepts
Proposed semi-supervised framework S4Crowd enhances crowd counting accuracy by leveraging unlabeled data and self-supervised losses.
Abstract
The content introduces the S4Crowd framework for semi-supervised crowd counting. It addresses the challenges of expensive crowd annotations by utilizing both labeled and unlabeled data. The framework includes self-supervised losses for crowd variation modeling, a Gated-Crowd-Recurrent-Unit for pseudo label generation, and a dynamic weighting strategy for balanced training. Extensive experiments on popular datasets show competitive performance.
Stats
Recent works achieved promising performance but relied on supervised paradigm with expensive crowd annotations.
Proposed S4Crowd framework leverages both unlabeled/labeled data for robust crowd counting.
Extensive experiments on four popular crowd counting datasets in semi-supervised settings.
Quotes
"Automatic crowd behavior analysis can help daily transportation statistics and planning."
"Our method achieved competitive performance in semi-supervised learning approaches on crowd counting datasets."