Your Cheat Sheet
The whole course on one page.
Here's the whole course, on a few screens. Pause, and take a screenshot.
- The whole course
- Pause and screenshot
- Keep it close
Which tool for which job?
Controlled scene, use classic tools. What is it, use a classifier with transfer learning. Where is it, use YOLO. Exact shape, use a segmenter. Across video, add a tracker. Read text, use OCR. Search images with words, use CLIP. And to prototype, use a multimodal language model.
| Job | Tool |
|---|---|
| Simple controlled scene | Classic OpenCV |
| What is it? | Transfer learning |
| Where is it? | YOLO |
| Exact shape | Segmentation |
| Across video | Tracker |
| Read text | OCR |
| Prototype | Multimodal LLM |
Which metric?
Classification, use recall on the important class. Detection, use MAP. Segmentation, use I o U or Dice. Tracking, use ID switches. And for the business, always report dollars saved.
| Task | Metric |
|---|---|
| Classification | Per-class recall |
| Detection | mAP |
| Segmentation | IoU, Dice |
| Tracking | ID switches |
| Business | Dollars saved |
The recipe
Here's the recipe. Fix the lighting. Collect real images. Start with pretrained models. Look at the failures. Price the mistakes. Export to the device. Run in shadow mode. Monitor, and retrain.
Summary
That's the entire course. One last lesson. Where do you go from here?
Key takeaways- Pick the right tool
- Measure what matters
- Follow the recipe