Test a New Camera Domain
A new camera has a different angle and sensor.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
A new camera has a different angle and sensor.
A practical question
The new garage camera has a lower angle and warmer lighting. The old model still produces confident occupancy labels, but those labels may no longer mean the same thing in this view.
The useful idea
Collect a labeled sample from the new camera before relying on it. Compare input appearance, error types, and calibration of scores with the original cameras. Make sure the parking-space geometry and camera crop are set correctly; a model cannot fix a box drawn over the wrong bay.
Work through the example
Run a limited review period and record disagreements. If performance falls, decide whether camera calibration, capture setup, additional training data, or a new threshold is justified. Evaluate changes on a further holdout. A newly installed camera can change viewpoint, lens distortion, exposure, and the visible shape of parking spaces. Test images from it before reusing old regions or confidence thresholds. Overlay the existing space map to see whether lines still land on the intended stalls. If not, recalibrate and evaluate again with labeled examples from the new view. A model trained on cars may still recognize cars while the occupancy geometry is wrong. Now check the idea on a concrete example: Test held-out images from the new camera before reusing old regions and thresholds. A possible trap is worth naming: A strong result on the original camera does not guarantee transfer. This supports a careful conclusion: Domain shift changes geometry, colors, resolution, or background from the training data. Place the old space map over a new-camera frame. If its outlines miss the stalls, recalibrate that geometry before blaming the car detector. Then test occupancy under the new angle with independent labels. A good detector cannot compensate for a stall region drawn over the driving lane.
Limits and next step
Accuracy alone isn't the complete user experience. Next, we'll measure how long the camera-to-result path takes.
Worked example
Test held-out images from the new camera before reusing old regions and thresholds.
Quick check: test your understanding
Try it yourself: Consider this case: Test held-out images from the new camera before reusing old regions and thresholds. Explain how you would check the result. In your answer, address this limitation: A strong result on the original camera does not guarantee transfer.
Check the reasoning
Domain shift changes geometry, colors, resolution, or background from the training data. A strong result on the original camera does not guarantee transfer.
What is the main idea behind test a new camera domain?
Domain shift changes geometry, colors, resolution, or background from the training data.
How is it used in the example?
Test held-out images from the new camera before reusing old regions and thresholds.
What limitation should be checked?
A strong result on the original camera does not guarantee transfer.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 72 transcript300 words · Narration ready for the upcoming videoOpen transcript
The new garage camera has a lower angle and warmer lighting. The old model still produces confident occupancy labels, but those labels may no longer mean the same thing in this view.
Collect a labeled sample from the new camera before relying on it. Compare input appearance, error types, and calibration of scores with the original cameras. Make sure the parking-space geometry and camera crop are set correctly; a model cannot fix a box drawn over the wrong bay.
Run a limited review period and record disagreements. If performance falls, decide whether camera calibration, capture setup, additional training data, or a new threshold is justified. Evaluate changes on a further holdout. A newly installed camera can change viewpoint, lens distortion, exposure, and the visible shape of parking spaces. Test images from it before reusing old regions or confidence thresholds. Overlay the existing space map to see whether lines still land on the intended stalls. If not, recalibrate and evaluate again with labeled examples from the new view. A model trained on cars may still recognize cars while the occupancy geometry is wrong. Now check the idea on a concrete example: Test held-out images from the new camera before reusing old regions and thresholds. A possible trap is worth naming: A strong result on the original camera does not guarantee transfer. This supports a careful conclusion: Domain shift changes geometry, colors, resolution, or background from the training data. Place the old space map over a new-camera frame. If its outlines miss the stalls, recalibrate that geometry before blaming the car detector. Then test occupancy under the new angle with independent labels. A good detector cannot compensate for a stall region drawn over the driving lane.
Accuracy alone isn't the complete user experience. Next, we'll measure how long the camera-to-result path takes.