Anchor Box

The mathematical mechanism that refines anchor box coordinates. Instead of predicting absolute coordinates, the network predicts offsets (tx, ty, tw, th) relative to the anchor box center and dimensions.

• Log-space scaling: Height and width offsets are predicted in log space to handle exponential size variations
• Smooth L1 loss: The standard regression loss function that is less sensitive to outliers than L2 loss
• Variance encoding: Some frameworks use variance-weighted offsets to normalize the regression targets across different anchor scales

The primary metric for matching anchor boxes to ground truth annotations during training. An anchor is considered a positive sample if its IoU with any ground truth box exceeds a threshold (typically 0.7).

• IoU thresholds: Higher thresholds (0.7+) produce higher quality proposals but fewer positive samples
• Anchor matching strategy: Each ground truth box is matched to the anchor with the highest IoU, ensuring every object gets at least one anchor
• Negative mining: Anchors with IoU < 0.3 are labeled as background; those in between are ignored during training

The fully convolutional sub-network in Faster R-CNN that generates region proposals using anchor boxes as reference templates. The RPN slides a small network over the convolutional feature map and at each sliding-window location, predicts k anchor boxes with different scales and aspect ratios.

• Classification head: Outputs 2k scores estimating the probability of object vs. background for each anchor
• Regression head: Outputs 4k coordinates encoding the bounding box refinement offsets
• Multi-scale detection: Anchors at different feature pyramid levels handle objects of varying sizes, from small nodules to large tumors

A modern paradigm that eliminates the need for hand-crafted anchor boxes entirely. Instead of pre-defining reference boxes, these methods directly predict keypoints (e.g., object centers, corners) or per-pixel bounding boxes.

• CenterNet: Predicts object centers as heatmap peaks and regresses width/height directly
• FCOS: Treats each pixel as a training sample and predicts distances to the four sides of the bounding box
• Advantages: Eliminates anchor-related hyperparameters (scales, ratios, IoU thresholds) and reduces the number of candidate detections requiring NMS

The backbone architecture that assigns anchor boxes to different resolution levels based on their scale. FPN builds a pyramidal hierarchy of feature maps where larger anchors are assigned to coarser, semantically stronger feature maps and smaller anchors to finer, high-resolution maps.

• Level assignment formula: An anchor of width w and height h is assigned to level P_k where k = floor(k_0 + log_2(sqrt(w*h)/224))
• Medical imaging benefit: Enables detection of abnormalities spanning orders of magnitude in size, from micro-calcifications to large masses
• Shared heads: All pyramid levels share the same classification and regression sub-networks

The post-processing step that filters the dense set of anchor-derived proposals down to a sparse set of final detections. After regression, thousands of overlapping boxes may predict the same lesion; NMS selects the one with the highest confidence score and suppresses others with high IoU overlap.

• IoU threshold for suppression: Typically 0.5; lower values are more aggressive at removing duplicates
• Soft-NMS: Instead of zeroing out overlapping boxes, decays their scores proportionally to IoU, preserving nearby distinct objects
• Class-specific NMS: Applied independently per class to avoid suppressing different pathologies that happen to overlap spatially