Lec10 Object Detection I

Detection = Localization + Classification

• Localization: Regression Problem

  • of parameters of box = 4 (x, y, w, h)

  • L2 Loss
• Classification
  • XE Loss
• Multi-task Learning, Multi-task Loss
• Base Network: feature extraction
• 일반적으로 Pre-trained on ImageNet 모델 사용

Object Detection for Multiple objects

• Various shapes of candidate boxes
• Too many candidates (bbox or anchor box)
• Inefficient to search → Selective Search

Selective Search (Region Proposal)

• Check only whether the candidate box has an object or not (only check promising boxes)
• After filtering out nonpromising boxes, perform detection (localization + classifcation)

Intersection over Union (IOU)

• IOU = Intersection / Union
• 보통 IOU ≥ 0.5 → Correct detection
• Problem: no intersection → IOU=0
  • We cannot differentiate → Gradient descent 불가능

Non-maximum Suppression (NMS)