Cascade R-CNN

Finding an object in a photo is not enough; a detector also needs to draw a box that fits it closely. Cascade R-CNN improves this box-fitting process by checking each candidate object region several times, with each check demanding greater precision than the last.

Cascade R-CNN is a two-stage object detector built on the same broad idea as Faster R-CNN. First, a region proposal network identifies image areas likely to contain objects. Those areas then pass through a sequence of detection “heads.” Each head classifies the object and adjusts its bounding box, but it is trained with a stricter Intersection over Union (IoU) threshold than the one before it.

• An early stage accepts rough matches, such as boxes with IoU of 0.5.
• A middle stage learns to improve moderately accurate boxes, such as IoU 0.6.
• A later stage focuses on highly accurate boxes, such as IoU 0.7.

Think of it as measuring and trimming a picture frame: the first pass gets near the edges, while later passes make fine corrections. A single detector trained only for strict overlap struggles because it sees too few good training examples. A detector trained only for loose overlap finds objects but produces imprecise boxes. The cascade solves both problems: earlier stages create better candidates for later, more selective stages.

This matters when localization quality is important, including pedestrian detection for autonomous vehicles, identifying defects on a production line, and locating lesions in medical scans. It is especially strong under evaluation metrics such as COCO’s average precision, which reward boxes that align closely with the true object boundary. Cascade R-CNN is widely available in the MMDetection framework, where it can be paired with feature pyramid networks and Mask R-CNN-style segmentation heads. The trade-off is extra computation: each proposal must pass through multiple refinement stages, but the result is more reliable, tightly placed detections.