Focal Loss for Dense Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Two-stage object detectors, like R-CNN, achieve high accuracy by classifying sparse candidate locations.
• One-stage detectors, using dense sampling, offer speed and simplicity but historically lag in accuracy.
• The performance gap is attributed to challenges in training dense detectors.

Purpose of the Study:

• Investigate the reasons behind the lower accuracy of one-stage object detectors.
• Propose a novel loss function to overcome training challenges in dense object detection.
• Develop a high-accuracy, efficient one-stage object detection system.

Main Methods:

• Identified extreme foreground-background class imbalance as the primary cause of poor performance in dense detectors.
• Introduced a novel Focal Loss function to reshape the standard cross-entropy loss.
• Designed and trained RetinaNet, a simple dense object detector, using the Focal Loss.

Main Results:

• The Focal Loss effectively down-weights easily classified examples, focusing training on hard examples.
• RetinaNet, trained with Focal Loss, achieved state-of-the-art accuracy surpassing existing two-stage detectors.
• The proposed method maintains the speed advantage of one-stage detectors.

Conclusions:

• Extreme class imbalance is a critical issue in training dense object detectors.
• Focal Loss provides an effective solution to mitigate class imbalance, enhancing detector performance.
• RetinaNet demonstrates that one-stage detectors can achieve superior accuracy and speed with appropriate loss functions.