Anchor Box
Predefined bounding boxes of various sizes and aspect ratios that serve as starting points for object detection.
Anchor boxes are predefined bounding box templates in object detection models – modern anchor-free methods like DETR and FCOS eliminate them.
Explanation
Object detection models like Faster R-CNN, SSD, and earlier YOLO versions use anchor boxes as "proposals" that the model refines. Anchor-free methods (FCOS, newer YOLO) eliminate this step.
Marketing Relevance
Understanding anchor boxes is essential for tuning object detection models and choosing between anchor-based and anchor-free architectures.
Common Pitfalls
Wrong anchor sizes lead to poor detection. Too many anchors increase compute. K-means clustering on training data helps selection.
Origin & History
Faster R-CNN (Ren et al., 2015) introduced anchor boxes. SSD (2016) and YOLOv2 (2017) adopted the concept. The trend since 2020 is toward anchor-free detection (FCOS, CenterNet, DETR).
Comparisons & Differences
Anchor Box vs. Anchor-Free Detection
Anchor-based methods need predefined box templates. Anchor-free methods (FCOS, CenterNet) predict object centers and sizes directly.