Generalized Focal Loss: Towards Efficient Representation Learning for Dense Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection is crucial in computer vision, with dense detectors gaining popularity for their efficiency.
• Current dense detectors face challenges in representation and optimization for quality estimation, classification, and localization.
• Existing methods often use inflexible distributions and inconsistent practices between training and inference.

Purpose of the Study:

• To address limitations in existing dense object detection methods.
• To propose novel representations for quality estimation, classification, and localization.
• To develop an optimized loss function for improved detection accuracy and efficiency.

Main Methods:

• Introduced new representations for quality estimation, classification, and localization by merging quality estimation into class prediction and using vector representations for box locations.
• Developed Generalized Focal Loss (GFocal) to handle continuous labels arising from new representations, generalizing the original Focal Loss.
• Constructed NanoDet, a lightweight detector for mobile settings, based on GFocal.

Main Results:

• The proposed representations eliminate inconsistency risks and accurately depict flexible distributions.
• GFocal successfully optimizes continuous labels, improving detection performance.
• NanoDet achieved 1.8 AP higher, 2x faster, and 6x smaller than scaled YoloV4-Tiny on mobile settings.

Conclusions:

• The novel representations and GFocal significantly enhance dense object detection.
• The method maintains efficiency in both training and inference.
• GFocal enables the development of highly efficient and accurate lightweight detectors like NanoDet for mobile applications.