Localization¶

Image → Bounding box

1. Train classification model
2. Attach new fully-connected "regression head"
3. Train regression head with regression loss
4. At inference, use both heads

Regression Head Approaches¶

• Class agnostic: one box in total
• Class specific: one box per class
  • more intuitive
  • works for multiple object localization; For eg: represent human pose with \(k\) joints

Regression Head Position¶

Sliding Window¶

Naive¶

• Run classification + regression head at multiple location on high resolution network
• Combine classifier and regressor predictions across all scales for final prediction

Efficient¶

Convert FC layers into conv layers

Train
Inference (larger image)

Advantage: Extra compute only for extra pixels

Comments