Cascade R-CNN is a multi-stage object detection architecture designed to improve detection quality by addressing issues of overfitting and quality mismatch between detectors and hypotheses. The method uses a sequence of detectors trained with increasing IoU thresholds, each stage progressively refining the hypotheses to better match the detector's quality. This approach ensures that all detectors have a balanced set of positive examples, reducing overfitting and improving detection performance. The same cascade procedure is applied during inference, allowing the detector to better match the quality of the input hypotheses. The Cascade R-CNN is shown to outperform existing single-model detectors on the COCO dataset, achieving consistent gains across different detector architectures. The architecture is simple to implement and end-to-end trainable, with results demonstrating its effectiveness in improving detection accuracy. The method is applicable to various object detection frameworks and has been validated through extensive experiments, showing significant improvements in detection performance across different IoU thresholds. The Cascade R-CNN also demonstrates strong generalization capabilities, performing well on multiple baseline detectors and achieving high accuracy on the COCO dataset. The architecture is efficient in terms of computational resources and has been shown to outperform state-of-the-art detectors in various evaluation metrics.Cascade R-CNN is a multi-stage object detection architecture designed to improve detection quality by addressing issues of overfitting and quality mismatch between detectors and hypotheses. The method uses a sequence of detectors trained with increasing IoU thresholds, each stage progressively refining the hypotheses to better match the detector's quality. This approach ensures that all detectors have a balanced set of positive examples, reducing overfitting and improving detection performance. The same cascade procedure is applied during inference, allowing the detector to better match the quality of the input hypotheses. The Cascade R-CNN is shown to outperform existing single-model detectors on the COCO dataset, achieving consistent gains across different detector architectures. The architecture is simple to implement and end-to-end trainable, with results demonstrating its effectiveness in improving detection accuracy. The method is applicable to various object detection frameworks and has been validated through extensive experiments, showing significant improvements in detection performance across different IoU thresholds. The Cascade R-CNN also demonstrates strong generalization capabilities, performing well on multiple baseline detectors and achieving high accuracy on the COCO dataset. The architecture is efficient in terms of computational resources and has been shown to outperform state-of-the-art detectors in various evaluation metrics.