Skip to main content

Object Detection

This module introduces the core ideas behind object detection and the main model families used in modern computer vision.

Core concept​

Object detection answers two questions:

  • What object is present?
  • Where is it located?

The output is usually:

  • a class label
  • a bounding box

Common evaluation terms:

  • IoU: overlap between predicted and ground-truth boxes
  • Precision/Recall
  • mAP: standard detection metric

Key model notes​

YOLO​

  • One-stage detector designed for speed
  • Predicts boxes and classes in a single pass
  • Common choice for real-time systems
  • Paper: https://arxiv.org/pdf/1506.02640
  • Hugging Face: placeholder
  • PPT: placeholder
  • Notebook: placeholder

DETR​

  • Transformer-based detector
  • Uses object queries instead of hand-crafted anchors
  • Removes the need for NMS in the standard formulation
  • Paper: https://arxiv.org/pdf/2005.12872
  • Hugging Face: placeholder
  • PPT: placeholder
  • Notebook: placeholder

DINO​

  • Improves DETR with denoising and better training behavior
  • Stronger convergence and more stable optimization
  • Paper: https://arxiv.org/pdf/2203.03605
  • Hugging Face: placeholder
  • PPT: placeholder
  • Notebook: placeholder

Grounding DINO​

  • Open-vocabulary detector
  • Connects text prompts with visual regions
  • Useful for language-guided detection and zero-shot settings
  • Paper: https://arxiv.org/pdf/2303.05499
  • Hugging Face: placeholder
  • PPT: placeholder
  • Notebook: placeholder

Shared resources​

  • Data (zip file): placeholder

Suggested study flow​

  1. Review object detection basics
  2. Study YOLO
  3. Study DETR
  4. Compare DINO and Grounding DINO
  5. Practice with notebooks and small examples