Temel Kavramlar
Proposing Shifted Autoencoders (SAE) to refine initial point annotations, improving object counting accuracy.
Özet
The content discusses the challenges of inconsistency in point annotations for object counting tasks and introduces the Shifted Autoencoders (SAE) method to address this issue. The SAE leverages general positional knowledge to restore shifted annotations, resulting in specific-offset-noise-free restoration. Extensive experiments across various datasets demonstrate the effectiveness of using refined annotations for training counting models.
Directory:
Abstract
Object counting relies on 2D point annotations.
Annotation inconsistency affects model training.
Introduction
Object counting methods categorized into localization-based and density-map-based approaches.
Methodology
SAE network architecture and training objective.
Experiments
Evaluation on eleven datasets across three domains: crowd counting, remote sensing object counting, and cell counting.
Conclusion
Proposed SAE improves consistency in point annotations for enhanced object counting accuracy.
İstatistikler
Remarkably, the proposed SAE helps set new records on nine datasets.