[Paper Review] PyramidBox++: High Performance Detector for Finding Tiny Face

Zhihang Li, Xu Tang+3 more|arXiv (Cornell University)|Mar 31, 2019

Face recognition and analysisComputer Science44 references37 citations

TL;DR

PyramidBox++ enhances tiny-face detection on WIDER FACE using balanced data sampling, dual-pyramid anchors, a dense context module, and multi-task training to achieve state-of-the-art results on the hard set.

ABSTRACT

With the rapid development of deep convolutional neural network, face detection has made great progress in recent years. WIDER FACE dataset, as a main benchmark, contributes greatly to this area. A large amount of methods have been put forward where PyramidBox designs an effective data augmentation strategy (Data-anchor-sampling) and context-based module for face detector. In this report, we improve each part to further boost the performance, including Balanced-data-anchor-sampling, Dual-PyramidAnchors and Dense Context Module. Specifically, Balanced-data-anchor-sampling obtains more uniform sampling of faces with different sizes. Dual-PyramidAnchors facilitate feature learning by introducing progressive anchor loss. Dense Context Module with dense connection not only enlarges receptive filed, but also passes information efficiently. Integrating these techniques, PyramidBox++ is constructed and achieves state-of-the-art performance in hard set.

Motivation & Objective

• Motivate robust face detection in unconstrained environments with large scale variance and tiny faces.
• Improve PyramidBox by addressing data sampling imbalance and context utilization.
• Leverage multi-task learning to provide additional supervision and regularization.

Proposed method

• Introduce Balanced-data-anchor-sampling (BDAS) to achieve uniform face scale sampling.
• Propose Dual-PyramidAnchors to enable progressive contextual supervision near the backbone.
• Incorporate a Dense Context Module with dense connections to enlarge receptive field and improve information flow.
• Employ multi-task training including segmentation and anchor-free detection to provide extra supervision.

Experimental results

Research questions

• RQ1 How can sampling balance across face scales improve tiny-face recall on WIDER FACE?
• RQ2 Does incorporating dual-pyramid anchors and dense context improve small-face localization compared to prior PyramidBox?
• RQ3 What is the impact of segmentation and anchor-free tasks on overall detection performance for tiny faces?

Key findings

• On WIDER FACE, PyramidBox++ achieves AP of 96.5% (Easy), 95.9% (Medium), 91.2% (Hard) on validation and 95.6% (Easy), 95.2% (Medium), 90.9% (Hard) on testing.
• The method outperforms all approaches on the hard subset, indicating strong tiny-face detection performance.
• BDAS leads to a more uniform sampling of face scales, improving small-face recall.
• Dense Context Module and Dual-PyramidAnchors contribute to better feature learning near the backbone without extra runtime cost.
• Multi-task training with segmentation and anchor-free branches provides additional supervision that enhances detection performance.