Computer Vision Roadmap

Teaching machines to understand images and video.

Topics

Core tasks

• Image Classification — what is in this image?
• Object Detection — where are the objects? (bounding boxes)
• Image Segmentation — pixel-level classification
• Image Generation — creating new images (GANs, diffusion)

Motion and tracking

• Multi-Object Tracking — track objects across video frames with persistent IDs
• Optical Flow — pixel-level motion estimation between consecutive frames
• Video Understanding — temporal analysis: action recognition, anomaly detection

3D and depth

• 3D Vision and Depth — stereo vision, monocular depth, point clouds, SfM
• Tutorial - Visual SLAM Concepts — simultaneous localization and mapping

Human understanding

• Pose Estimation — detect body keypoints, skeleton-based action recognition

Techniques

• Convolutional Neural Networks — the architecture for vision
• Transfer Learning — pretrained models (ResNet, EfficientNet, ViT)
• Data Augmentation — expand training data artificially

Learning order

Phase 1: Foundations
  1. Convolutional Neural Networks (how vision models work)
  2. Image Classification (the "hello world")
  3. Transfer Learning (pretrained models)

Phase 2: Spatial tasks
  4. Object Detection (bounding boxes)
  5. Image Segmentation (pixel-level)
  6. Pose Estimation (body keypoints)

Phase 3: Temporal tasks
  7. Optical Flow (pixel motion)
  8. Multi-Object Tracking (identity across frames)
  9. Video Understanding (what's happening over time)

Phase 4: 3D understanding
  10. 3D Vision and Depth (depth estimation, point clouds)
  11. Visual SLAM Concepts (localization + mapping)

Phase 5: Applied
  12. Tutorial - Object Tracking Pipeline (build a tracker)
  13. Tutorial - Aerial Image Analysis (satellite/drone imagery)
  14. Case Study - CV Pipeline Design (design judgment)

The modern workflow

1. Pick a pretrained model (torchvision, timm, huggingface)
2. Replace the classification head
3. Apply data augmentation
4. Fine-tune on your data
5. Evaluate and iterate

You almost never train a vision model from scratch.

Tutorials and applied

• Tutorial - Object Tracking Pipeline — build a multi-object tracker from scratch
• Tutorial - Aerial Image Analysis — satellite imagery, change detection, NDVI
• Tutorial - Visual SLAM Concepts — visual odometry and mapping

Design judgment

• Case Study - CV Pipeline Design — three real scenarios: drone tracking, satellite change detection, indoor activity monitoring

Key libraries

• torchvision — datasets, models, transforms
• timm — huge collection of pretrained models
• albumentations — fast image augmentation
• ultralytics — YOLO for detection and tracking
• opencv-python — classical CV, optical flow, feature matching
• open3d — point cloud processing and visualization
• rasterio — geospatial image loading (satellite data)
• mediapipe — pose estimation, face/hand detection (on-device)

Links

• Deep Learning Roadmap
• Convolutional Neural Networks
• Transfer Learning
• Reinforcement Learning Roadmap — RL for autonomous vision systems