Localization in the Crowd with Topological Constraints

We address the problem of crowd localization, i.e., prediction dots corresponding to people in a crowded scene. Due various challenges, localization method is prone spatial semantic errors, predicting multiple within same person or collapsing cluttered region. propose topological approach targeting these errors. introduce constraint that teaches model reason about arrangement dots. To enforce this constraint, we define persistence loss based on theory persistent homology. The compares topographic landscape likelihood map and topology ground truth. Topological reasoning improves quality algorithm especially near regions. On public benchmarks, our outperforms previous methods. Additionally, demonstrate potential improving performance counting task.