D-DL4CV-Lec15cb-NMS

Non-Max Suppression (NMS)

Introduction

NMS is a method aims to resolve the problem of multiple overlapping detections for the same object

What it does is comparing IoU between every two boxes and delete those with high IoU, which suggest they overlaps too much

Steps

1. Sort all detection boxes by confidence score (highest to lowest)
2. Select the next highest-scoring box from the remaining unprocessed boxes
3. Add selected box to final output
4. Remove all remaining boxes with $\text{IoU}>\text{threshold}$ (e.g. 0.7) with the selected box
5. If any boxes remain unprocessed, GOTO 2

Problem

When objects are highly overlapping, like swarms of people at MRT in peak hours, we may delete good boxes