Cascade R-CNN: High Quality Object Detection and Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Intersection over Union (IoU) threshold is crucial for defining positive/negative samples in object detection.
• Commonly used IoU thresholds (e.g., 0.5) can lead to noisy detections, while higher thresholds often degrade performance.
• Existing methods face a paradox where increasing IoU thresholds for higher quality detections leads to overfitting and inference-time mismatches.

Purpose of the Study:

• To propose a novel multi-stage object detection architecture, Cascade R-CNN, to address the paradox of high-quality detection.
• To improve detection performance by mitigating overfitting and resolving quality mismatches between detectors and hypotheses.
• To generalize the Cascade R-CNN architecture for instance segmentation tasks.

Main Methods:

• A sequential cascade of detectors is trained with progressively increasing IoU thresholds.
• The output of each detector is used as the training set for the subsequent detector in the cascade.
• The same cascaded structure is applied during inference to ensure quality consistency.

Main Results:

• The Cascade R-CNN architecture achieves state-of-the-art performance on the COCO dataset.
• Significant improvements in high-quality detection are observed across various datasets (VOC, KITTI, CityPerson, WiderFace).
• The architecture, when generalized to instance segmentation, shows notable improvements over Mask R-CNN.

Conclusions:

• Cascade R-CNN effectively resolves the paradox of high-quality object detection by addressing overfitting and inference-time mismatches.
• The proposed sequential training and inference strategy enhances detection quality and robustness.
• The architecture offers a promising direction for both object detection and instance segmentation.