[Paper Review] Face Detection Using Improved Faster RCNN

Changzheng Zhang, Xiang Xu+1 more|arXiv (Cornell University)|Feb 6, 2018

Face recognition and analysisComputer Science30 references64 citations

TL;DR

This paper presents FDNet1.0, an improved Faster RCNN-based face detector that uses multi-scale training/testing, a lighter RCNN design, inference tricks, and a vote-based ensemble to achieve top results on WIDER FACE.

ABSTRACT

Faster RCNN has achieved great success for generic object detection including PASCAL object detection and MS COCO object detection. In this report, we propose a detailed designed Faster RCNN method named FDNet1.0 for face detection. Several techniques were employed including multi-scale training, multi-scale testing, light-designed RCNN, some tricks for inference and a vote-based ensemble method. Our method achieves two 1th places and one 2nd place in three tasks over WIDER FACE validation dataset (easy set, medium set, hard set).

Motivation & Objective

• Motivate the use of Faster RCNN for face detection and push performance on challenging face datasets.
• Develop a streamlined, lightweight Faster RCNN variant suitable for faces (FDNet1.0).
• Incorporate multi-scale training/testing and ensemble techniques to boost accuracy.
• Demonstrate effectiveness through competitive results on the WIDER FACE dataset.

Proposed method

• Adopt an enhanced Faster RCNN framework named FDNet1.0 for face detection.
• Apply multi-scale training and multi-scale testing to improve detection across scales.
• Design a light-weight Faster RCNN variant to reduce computation while maintaining accuracy.
• Incorporate inference tricks to optimize performance during deployment.
• Introduce a vote-based ensemble method to combine predictions from multiple models.

Experimental results

Research questions

• RQ1 Can multi-scale training and testing improve face detection accuracy on diverse scales?
• RQ2 Does a lightweight Faster RCNN design (FDNet1.0) retain high accuracy for faces while reducing computation?
• RQ3 Can inference tricks and a vote-based ensemble further boost detection performance on WIDER FACE?

Key findings

• FDNet1.0 achieves top performance (two 1st places) and one 2nd place on the WIDER FACE validation set across easy, medium, and hard subsets.