[Paper Review] TinaFace: Strong but Simple Baseline for Face Detection

Yanjia Zhu, Hongxiang Cai+3 more|arXiv (Cornell University)|Nov 26, 2020

Advanced Neural Network ApplicationsComputer Science62 references68 citations

TL;DR

TinaFace treats face detection as a one-class generic object detection problem and achieves state-of-the-art results on WIDER FACE with a strong but simple one-stage baseline based on RetinaNet. It reaches 92.1% AP on hard, 92.4% AP with test-time augmentation.

ABSTRACT

Face detection has received intensive attention in recent years. Many works present lots of special methods for face detection from different perspectives like model architecture, data augmentation, label assignment and etc., which make the overall algorithm and system become more and more complex. In this paper, we point out that extbf{there is no gap between face detection and generic object detection}. Then we provide a strong but simple baseline method to deal with face detection named TinaFace. We use ResNet-50 \cite{he2016deep} as backbone, and all modules and techniques in TinaFace are constructed on existing modules, easily implemented and based on generic object detection. On the hard test set of the most popular and challenging face detection benchmark WIDER FACE \cite{yang2016wider}, with single-model and single-scale, our TinaFace achieves 92.1\% average precision (AP), which exceeds most of the recent face detectors with larger backbone. And after using test time augmentation (TTA), our TinaFace outperforms the current state-of-the-art method and achieves 92.4\% AP. The code will be available at \url{https://github.com/Media-Smart/vedadet}.

Motivation & Objective

• Demonstrate that face detection can be effectively approached as a one-class generic object detection problem.
• Provide a strong, simple baseline (TinaFace) built on standard detection modules.
• Show that TinaFace achieves state-of-the-art or competitive results on the WIDER FACE hard subset with a ResNet-50 backbone.

Proposed method

• Use ResNet-50 with a 6-level FPN backbone as a common, fair baseline.
• Incorporate an inception module to enhance multi-scale features.
• Add an IoU-aware head to predict box IoU for reweighting confidence.
• Use DIoU loss for bounding-box regression to better handle small objects.
• Replace BatchNorm with GroupNorm for stable training with small batch sizes.
• Optionally employ test-time augmentation (TTA) to boost performance.

Experimental results

Research questions

• RQ1 Can face detection be effectively solved using standard generic object detection techniques without specialized face-specific components?
• RQ2 What performance can be achieved on WIDER FACE hard subset using a simple, single-scale, single-model pipeline?
• RQ3 Does incorporating IoU-awareness and distance-IoU loss improve localization and overall AP, especially for small faces?

Key findings

• With a single-scale, single-model setup and ResNet-50 backbone, TinaFace reaches 92.1% AP on the hard subset of WIDER FACE.
• With test-time augmentation, TinaFace attains 92.4% AP on the hard subset, surpassing current state-of-the-art methods.
• A strong baseline built on generic detection modules can outperform many recent face detectors that use larger backbones.
• DIoU loss and IoU-aware head contribute to improved localization and reduced false positives on small faces.
• Replacing BatchNorm with GroupNorm stabilizes training under small-batch conditions.
• Table results show TinaFace achieving 0.963 Easy, 0.957 Medium, 0.930 Hard on validation and 0.952 Easy, 0.947 Medium, 0.921 Hard on test with ResNet-50 (no-TTA).