Faster R-CNN: Bounding Box Regression Targets

Advanced

The problem, from first principles

The region proposal network refines each anchor box by predicting four offsets: a centre shift scaled by the box size, and a log scale change for width and height. These are the targets the network regresses toward.

From theory to code

Implement box_targets(anchor, gt), returning (tx, ty, tw, th) from the anchor to the ground truth.

Constraints

Widths and heights are positive.

Hints

Hint 1

Convert both boxes to center and size. Take center differences divided by the anchor size, and log ratios of sizes.

Keywords

research-paperscomputer-visiondetectionobject-detection

Python

Run your code to see results

No results yet

Run your code to check it against the examples, or submit to run every test.