Start with a Feature-Based Baseline
The team needs a transparent damaged-sign classifier.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
The team needs a transparent damaged-sign classifier.
A practical question
A maintenance team wants to classify photos as damaged or undamaged signs. Before choosing a large vision model, build a simple baseline from measurable features: perhaps edge shape, color, or texture in a standardized sign crop.
The useful idea
Use only features available at prediction time. Train a straightforward classifier and inspect the images it gets wrong. A simple method might work when the camera and sign type are consistent; it may fail when paint colors and backgrounds vary. Those failures tell us what a more complex method would need to handle.
Work through the example
Compare the baseline on sign photos from the same camera and a held-out camera. If accuracy collapses on the new view, the data and preprocessing deserve attention before adding model complexity. Start with standardized crops of parking signs and simple features for edge shape, color, and texture. Compare an intact sign with a bent one from the same camera, then test both under a held-out camera. If performance collapses, investigate differences in viewpoint, lighting, and crop before choosing a larger model. Keep the feature list and dataset split so a later method is compared fairly. A baseline is useful precisely because its limitations are easy to inspect. Use this case to test your understanding: Test simple shape, color, and texture features on held-out camera views. Here is the limitation to watch for: Handcrafted features may miss subtle patterns but make useful comparisons. The answer follows from that evidence: A feature-based baseline measures interpretable color, texture, or shape summaries before classification. A damaged-sign baseline can reveal that edge features work on one sign type but fail when a camera crops the bent corner. Inspect that crop before replacing the classifier. The source image and preprocessing rule may matter more than the choice of model.
Limits and next step
Keep the feature definition and split. Next, we'll use nearby examples as a direct classification rule and examine when visual similarity helps or hurts.
Worked example
Test simple shape, color, and texture features on held-out camera views.
Quick check: test your understanding
Try it yourself: Consider this case: Test simple shape, color, and texture features on held-out camera views. Explain how you would check the result. In your answer, address this limitation: Handcrafted features may miss subtle patterns but make useful comparisons.
Check the reasoning
A feature-based baseline measures interpretable color, texture, or shape summaries before classification. Handcrafted features may miss subtle patterns but make useful comparisons.
What is the main idea behind start with a feature-based baseline?
A feature-based baseline measures interpretable color, texture, or shape summaries before classification.
How is it used in the example?
Test simple shape, color, and texture features on held-out camera views.
What limitation should be checked?
Handcrafted features may miss subtle patterns but make useful comparisons.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 49 transcript316 words · Narration ready for the upcoming videoOpen transcript
A maintenance team wants to classify photos as damaged or undamaged signs. Before choosing a large vision model, build a simple baseline from measurable features: perhaps edge shape, color, or texture in a standardized sign crop.
Use only features available at prediction time. Train a straightforward classifier and inspect the images it gets wrong. A simple method might work when the camera and sign type are consistent; it may fail when paint colors and backgrounds vary. Those failures tell us what a more complex method would need to handle.
Compare the baseline on sign photos from the same camera and a held-out camera. If accuracy collapses on the new view, the data and preprocessing deserve attention before adding model complexity. Start with standardized crops of parking signs and simple features for edge shape, color, and texture. Compare an intact sign with a bent one from the same camera, then test both under a held-out camera. If performance collapses, investigate differences in viewpoint, lighting, and crop before choosing a larger model. Keep the feature list and dataset split so a later method is compared fairly. A baseline is useful precisely because its limitations are easy to inspect. Use this case to test your understanding: Test simple shape, color, and texture features on held-out camera views. Here is the limitation to watch for: Handcrafted features may miss subtle patterns but make useful comparisons. The answer follows from that evidence: A feature-based baseline measures interpretable color, texture, or shape summaries before classification. A damaged-sign baseline can reveal that edge features work on one sign type but fail when a camera crops the bent corner. Inspect that crop before replacing the classifier. The source image and preprocessing rule may matter more than the choice of model.
Keep the feature definition and split. Next, we'll use nearby examples as a direct classification rule and examine when visual similarity helps or hurts.