10.0 Learning Checklist: Computer Vision

Use this page as a printable checklist. If you need the full explanation, return to the Chapter 10 entry page.

Two-Hour First Pass

Time box	Do this	Stop when you can say
20 min	Read the output-granularity ladder	”Classification, detection, and segmentation differ by output.”
25 min	Run the pixel lab	”I can inspect size, channels, RGB, and grayscale.”
25 min	Skim 10.1 image basics	”Preprocessing changes the data the model sees.”
25 min	Skim classification, detection, segmentation roadmaps	”I know which metric belongs to which task.”
25 min	Read the debugging loop	”I should inspect data and labels before blaming architecture.”

Required Evidence

Evidence	Minimum version
opencv_demo.py or pixel_lab.py	image load or generated image, preprocessing, saved output
vision_dataset.md	data source, classes, sample count, annotation method, limitations
eval_results.md	accuracy/F1, mAP, IoU/Dice, OCR hit rate, or chosen metric
failure_cases.md	failed images, possible cause, fix direction
README.md	task goal, run command, input/output examples, scenario boundary

Quality Gates

Gate	Pass condition
Visual trace	Original, processed, prediction, and failure images are saved with matching filenames.
Annotation	Dataset notes define classes, boxes or masks, source, split, and known label uncertainty.
Metric fit	Accuracy/F1, mAP, IoU/Dice, or OCR hit rate matches the task output.
Real-world boundary	Report names lighting, angle, camera or source, latency, image size, and device limits.

Exit Questions

• Can you explain classification, detection, segmentation, and OCR by output shape?
• Can you show the original image, processed image, and prediction visualization?
• Can you explain why annotation quality affects metrics?
• Can you choose accuracy/F1, mAP, IoU, or Dice for the right task?
• Can you explain why a demo may fail on real images?

If the answer is yes, you can connect vision to multimodal work in Chapter 12.

Check reasoning and explanation

1. A passing answer maps the task to the right visual output: class label, bounding box, mask, OCR text, embedding, or video event.
2. The evidence should include a rendered visual artifact and one metric or qualitative error note.
3. A good self-check names one visual failure mode such as class confusion, missed objects, bad masks, lighting shift, domain shift, or weak annotation quality.

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Task Output

classification label, detection box, segmentation mask, OCR text, or video event

Artifacts

original image, processed image, prediction overlay, metrics file, and failure samples

Metric

accuracy/F1, mAP, IoU, Dice, latency, or scenario-specific review score

Failure Check

data quality, label error, preprocessing mismatch, threshold, or deployment constraint

Expected Output

a reproducible run folder with visual outputs and a short failure report