AP-Loss for Accurate One-Stage Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• One-stage object detectors face challenges with extreme foreground-background class imbalance during training.
• Traditional methods optimize classification and localization losses simultaneously, exacerbating imbalance issues due to numerous anchors.

Purpose of the Study:

• To propose a novel framework that replaces the classification task in one-stage object detectors with a ranking task.
• To address the foreground-background class imbalance issue inherent in object detection.

Main Methods:

• The study adopts the average-precision loss (AP-loss) for the ranking problem, which is non-differentiable and non-convex.
• A novel optimization algorithm is developed, combining perceptron learning's error-driven updates with deep network backpropagation.
• Theoretical and empirical analyses are conducted on the algorithm's convergence and computational complexity.

Main Results:

• The proposed framework significantly improves object detection performance by effectively addressing class imbalance.
• Experimental results show superior performance compared to existing AP-based optimization algorithms.
• The AP-loss based one-stage detectors achieve improved state-of-the-art results on standard benchmarks.

Conclusions:

• The novel ranking-based framework effectively mitigates class imbalance in one-stage object detectors.
• The developed optimization algorithm provides good convergence and computational efficiency.
• The framework demonstrates versatility and achieves state-of-the-art performance across various network architectures.