Bounding Box Regression

Rather than predicting absolute coordinates directly, regression models learn scale-invariant transformations between a prior box (anchor or proposal) and the ground truth. The network outputs offsets for the center coordinates $(t_x, t_y)$ and log-space scaling factors for the dimensions $(t_w, t_h)$. This normalization ensures the regression targets have zero mean and unit variance, stabilizing training. For example, a model refines a coarse anchor into a tight boundary around a pulmonary nodule by predicting a small $t_x$ shift and a $t_h$ scale factor.

The Smooth L1 loss is the standard objective for bounding box regression, combining the best properties of L1 and L2 losses. It is defined as $0.5x^2$ for $|x| < 1$ and $|x| - 0.5$ otherwise. This formulation is less sensitive to outliers than L2 loss, preventing exploding gradients from grossly mislocalized proposals, while remaining differentiable at zero. In medical imaging, this robustness is critical when a region proposal network initially places a box far from a small lesion.

Modern detectors often cascade multiple regression stages, where each stage is optimized for a specific Intersection over Union (IoU) threshold. A first stage regresses boxes to a 0.5 IoU, and a second stage refines those to 0.7. This sequential refinement, as seen in Cascade R-CNN, prevents overfitting that occurs when a single regressor is asked to handle proposals of drastically different quality. It produces higher-fidelity boundaries for irregularly shaped masses.

Standard regression losses optimize coordinate distances independently, which does not directly correlate with the final evaluation metric, IoU. Generalized IoU (GIoU) loss directly optimizes the overlap metric by incorporating a penalty term for the smallest enclosing box that covers both prediction and target. GIoU loss is defined as $1 - (IoU - \frac{|C \setminus (A \cup B)|}{|C|})$. This is particularly effective for improving the localization of large, diffuse findings like consolidations.

In class-agnostic regression, a single set of box offsets is predicted regardless of the object category, sharing localization knowledge across all classes. In class-specific regression, each class learns its own bounding box refinement parameters. For medical tasks, class-agnostic regression is often preferred for rare pathologies, as it allows the model to leverage the common shape characteristics of all lesions to improve localization, rather than overfitting to the few examples of a specific disease.

The regression targets are computed relative to matched anchor boxes or proposals. A positive match is typically assigned if the anchor has an IoU > 0.7 with a ground truth box, or is the highest-overlapping anchor. The regression layer then learns to transform this discrete, templated anchor into a continuous, precise location. Feature Pyramid Networks (FPN) assign anchors of different scales to different pyramid levels, ensuring that a small micro-calcification is regressed from a fine-grained feature map, not a coarse one.

A fully convolutional network that generates the initial region proposals that bounding box regression subsequently refines. The RPN slides a small network over a feature map, simultaneously predicting objectness scores and bounding box coordinates at each position using anchor boxes. This eliminates the need for external proposal algorithms like Selective Search, enabling end-to-end training in frameworks like Faster R-CNN.

• Predicts object bounds and objectness at each spatial location
• Uses anchor boxes as reference templates for multi-scale detection
• Shares convolutional features with the downstream detection network

The primary evaluation metric that quantifies the accuracy of a regressed bounding box by measuring its overlap with the ground truth annotation. Calculated as the area of overlap divided by the area of union, IoU produces a score between 0 and 1. During training, IoU thresholds determine positive vs. negative samples; during inference, they define true positives for mAP calculation.

• Formula: IoU = Area of Overlap / Area of Union
• Common thresholds: 0.5, 0.75, 0.5:0.95 for COCO-style evaluation
• Directly drives the regression loss function during training

The standard loss function for bounding box regression that combines the best properties of L1 loss (robustness to outliers) and L2 loss (smooth gradients near zero). It applies a squared term when the error is small and a linear term when the error is large, preventing exploding gradients from poorly localized proposals while maintaining precision for fine-tuning.

• Defined piecewise with a transition threshold (typically β=1.0)
• Less sensitive to outliers than pure L2 loss
• Used in Faster R-CNN, SSD, and many two-stage detectors

A post-processing algorithm that cleans up the output of bounding box regression by eliminating redundant, overlapping detections for the same object. NMS sorts all predicted boxes by confidence score, selects the highest-scoring box, and suppresses any remaining boxes with an IoU exceeding a threshold (typically 0.5). This ensures each lesion or abnormality is reported only once.

• Greedy algorithm: iteratively selects highest-confidence box
• Critical for reducing false positives in dense detection scenarios
• Variants include Soft-NMS and DIoU-NMS for improved performance

Predefined bounding box templates with various scales and aspect ratios that serve as reference priors for bounding box regression. Rather than predicting absolute coordinates, the network predicts offsets relative to anchor boxes, making the regression problem translation-invariant and easier to learn. Typical configurations include 3 scales × 3 aspect ratios at each sliding window position.

• Common aspect ratios: 1:1, 1:2, 2:1
• Scales adapted to object sizes in the target domain (e.g., small nodules vs. large tumors)
• Anchor-free alternatives like CenterNet and FCOS eliminate this hyperparameter

The definitive evaluation metric for object detection that summarizes bounding box regression quality across all classes and IoU thresholds. Average Precision (AP) computes the area under the precision-recall curve for a single class; mAP averages AP across all classes. Modern benchmarks like COCO evaluate mAP at multiple IoU thresholds (0.5:0.95) to assess localization accuracy at varying strictness levels.

• [email protected]: Traditional PASCAL VOC metric (lenient localization)
• [email protected]:0.95: COCO primary metric (rigorous localization)
• Directly reflects the quality of bounding box regression outputs