Group Sampling for Scale Invariant Face Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Current deep learning object detectors often utilize multi-scale feature maps for efficient multi-scale object detection.
• Techniques like Feature Pyramid Network (FPN) and Single Shot Detector (SSD) typically employ feature maps from various network layers to handle objects of different sizes.

Purpose of the Study:

• To investigate the factors influencing object detection performance across a wide range of scales.
• To propose a novel training strategy that enables effective detection of multi-scale objects using single-layer features.

Main Methods:

• Proposed a group sampling method that categorizes anchors by scale and ensures balanced positive and negative sample counts for each scale group during training.
• Evaluated the method using features from a single layer of the Feature Pyramid Network (FPN).

Main Results:

• Achieved state-of-the-art performance on face detection benchmarks, including FDDB and WIDER FACE.
• Demonstrated the effectiveness of the proposed group sampling method for improving detection across all scales.
• Showcased the adaptability of the approach to other datasets (e.g., COCO) and detection pipelines (e.g., YOLOv3, SSD, R-FCN).

Conclusions:

• The balance of training samples across different object scales is crucial for effective multi-scale detection, even with single-layer features.
• The proposed group sampling method offers a simple yet powerful solution for enhancing object detection performance.
• The approach is versatile and can be readily integrated into various object detection systems.