COCO-AD: Setting the Standard for Anomaly Detection with a Comprehensive Benchmark

Jiangning Zhang and the team address the current limitations of anomaly detection (AD) datasets by introducing COCO-AD. By extending COCO, they provide a large-scale, general-purpose dataset that enhances evaluation and encourages sustainable methods’ development.

• COCO-AD provides a platform for fair and extensive method evaluation.
• Novel metrics such as m\(F_1\)\(^{.2}_{.8}\) and mIoU\(^{.2}_{.8}\) offer a more practical evaluation of methods.
• The InvAD framework, based on GAN inversion, achieves high-quality feature reconstruction, improving the performance of reconstruction-based methods.
• Comprehensive experiments validate the effectiveness of InvAD’s components.

This research brings a significant contribution to the field of AD by offering a robust benchmark and inventive methodologies for feature reconstruction. COCO-AD’s introduction could revolutionize anomaly detection in various industries, from manufacturing to healthcare. Visit the arXiv abstract or the full PDF for more insights.