Anchor Box Matching for Object Detection

Before a convolutional detector can draw a single bounding box, it has already tiled the image with hundreds of candidate rectangles — anchor boxes — at every cell of its feature-map grid, in a handful of scales and aspect ratios. This simulator renders that grid in 3D exactly as a detector like SSD or Faster R-CNN's region-proposal network sees it: adjust the grid resolution and anchor scale, drag the two Intersection-over-Union thresholds, and drop a new random ground-truth box to watch, anchor by anchor, which shapes get accepted as positive training targets, which are dismissed as background negatives, and which fall into the ignored middle band — with the true governing IoU formula computed live for every single anchor.