DenseControl: Instance-Level Controllable Synthesis of Dense Crowd Image
Researchers have developed DenseControl, a new pipeline for generating dense crowd images with precise control over individual instance placement, scale, and attributes. The system addresses challenges in signal embedding and topological integrity by introducing an Isolated Object Embedding (IOE) map and an Implicit Scale Embedding (ISE) strategy, enhanced by a Position Shortcut mechanism. DenseControl achieves state-of-the-art results in crowd image synthesis and demonstrates practical utility in applications like data augmentation for crowd analysis, transfer learning, and weather generalization. AI
IMPACT Enables more realistic and controllable synthetic data generation for computer vision tasks.