Aligning Bounding Boxes with Human Preference

Overview

Through a user study with 123 participants, we discovered that humans prefer larger bounding boxes over smaller ones with the same IoU. We propose an asymmetric bounding box regression loss that encourages detectors to predict boxes better aligned with human preference.

Key Contributions

• Large-scale user study (N=123) on bounding box preferences
• Demonstrates human preference for larger boxes regardless of IoU
• Proposes asymmetric loss for human-aligned predictions
• Shows improved alignment over fixed scaling approaches

Links

• arXiv
• PDF
• Springer