Computer Vision Metrics: IoU, Precision-Recall & mAP by Hand

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Convert five boxes from one frame of Corvid Robotics' warehouse camera from corner format (x1, y1, x2, y2) to width, height, centre and area, and write one coordinate in normalised YOLO format.
• Calculate the overlap of six ground-truth and predicted boxes with MAX and MIN, clamp non-overlapping boxes to zero, and turn intersection and union into the IoU.
• Match 12 person detections on four frames to the 10 labelled persons, greedily by confidence: TP if IoU ≥ 0.5 and the ground truth is still free, otherwise FP; then count the missed persons (FN), precision and recall.
• Keep only the detections above seven confidence thresholds and calculate TP, precision, recall and F1 at each, to see the trade-off a deployment threshold makes.
• Walk down the ranked detections, accumulate TP and FP, calculate precision and recall after every box and draw the precision-recall curve as a scatter chart.
• Smooth the zig-zag PR curve into interpolated precision with MAX and MAXIFS and calculate the average precision of the person class twice: all-point (area) and 11-point (Pascal VOC 2007 style).
• Average the AP of person, forklift and pallet into mAP at every IoU threshold, compute the COCO-style mAP@[.5:.95], compare it with an instance-weighted average and chart how AP falls as the IoU threshold gets stricter.
• Read a class confusion matrix of matched detections, measure the pallet-forklift mix-up, and compare IoU and Dice on the pixel counts of five segmentation masks.