FCOS: A Simple and Strong Anchor-Free Object Detector

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection is crucial for computer vision, enabling instance-level recognition and downstream applications.
• One-stage object detection methods offer simpler designs and competitive performance compared to two-stage approaches.
• Current state-of-the-art detectors often rely on pre-defined anchor boxes, introducing computational complexity and sensitive hyperparameters.

Purpose of the Study:

• To propose a fully convolutional one-stage object detector (FCOS) that operates in a per-pixel prediction fashion.
• To develop an object detection framework that is both anchor-box free and proposal free.
• To simplify the object detection pipeline by removing anchor-related computations and hyperparameters.

Main Methods:

• Developed FCOS, a fully convolutional one-stage object detector.
• Implemented a per-pixel prediction approach, analogous to semantic segmentation.
• Eliminated the reliance on pre-defined anchor boxes and proposals, simplifying training and reducing hyperparameters.

Main Results:

• FCOS achieves improved detection accuracy with a significantly simpler and more flexible framework.
• The anchor-free design eliminates complex intersection over union (IoU) calculations during training.
• The detector avoids sensitive anchor-related hyperparameters, enhancing robustness.

Conclusions:

• The proposed FCOS framework offers a simpler and strong alternative for object detection.
• Anchor-free object detection significantly streamlines the detection pipeline.
• FCOS demonstrates competitive performance and flexibility for various instance-level computer vision tasks.