Accurate and Robust Object Detection via Selective Adversarial Learning With Constraints.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Network (ConvNet)-based object detectors excel on clean images but falter on corrupted data (noise, blur, adverse weather).
• This performance gap limits practicability in security-sensitive applications.
• Current robustness methods often require extra labeled data, image restoration, or degrade performance on clean images.

Purpose of the Study:

• To develop a universal training approach, Selective Adversarial Learning with Constraints (SALC), that simultaneously enhances object detector precision and robustness.
• To address the performance degradation on clean images often associated with adversarial training.

Main Methods:

• Proposed a unified formulation for adversarial samples in multitask adversarial learning to diversify training data.
• Introduced a batch local comparison strategy with two Batch Normalization (BN) branches to balance accuracy and robustness by analyzing model bias and BN statistics.
• Implemented task-aware ratio thresholds to manage subtask losses and prevent performance degradation.

Main Results:

• SALC achieved state-of-the-art results on both clean benchmarks (Pascal VOC, MS-COCO) and corruption benchmarks (Pascal VOC-C, MS-COCO-C).
• Demonstrated improved robustness against various image corruptions without compromising performance on clean images.
• Validated the approach's applicability to various detectors without additional labeled data, inference costs, or model parameters.

Conclusions:

• SALC offers a novel and effective solution for improving the robustness and precision of object detection networks.
• The method provides a practical and efficient way to enhance detector performance in real-world, challenging conditions.
• SALC represents a significant advancement in developing reliable computer vision systems.