Evaluate Tracks and Identity Switches
A tracker counts vehicles well but keeps swapping their names.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
A tracker counts vehicles well but keeps swapping their names.
A practical question
Our parking tracker finds nearly every car box, but a red car changes identity twice while crossing the lot. A detection score might look excellent and still miss this problem.
The useful idea
Compare each predicted track with annotated identities over time. Count missed detections, false tracks, and identity switches, and inspect when they happen. An identity switch can matter greatly if the application counts individual visits or follows a vehicle between bays.
Work through the example
Show one clean sequence and one difficult crossing. Mark the frame where an identity changes, then trace whether the cause was a missed detection, an ambiguous association, or a long occlusion. A single aggregate score should be paired with these concrete errors. Detecting every car in every frame still does not guarantee good tracking. When two cars cross, their labels can swap while the boxes remain accurate. Review the crossing as a sequence, not a collection of isolated screenshots. Note where the identity changed, whether it later recovered, and how that affects the question the garage cares about. For occupancy, a brief swap may matter less than for a long-term vehicle count, so evaluate the error against the intended use. Now check the idea on a concrete example: Review one identity switch after two cars cross, even if every car was detected. A possible trap is worth naming: A good detection score does not guarantee stable tracks. This supports a careful conclusion: Tracking metrics include detection coverage and identity consistency, which measure different errors. When the red car changes labels, ask whether the garage's count changes too. The same switch may be serious for trip histories and harmless for a single occupancy snapshot. Report the identity error and the operational consequence separately.
Limits and next step
Decide which errors affect the parking decision and set review rules accordingly. Next, we'll leave video tracking and represent a still image so we can search a collection for similar scenes.
Worked example
Review one identity switch after two cars cross, even if every car was detected.
Quick check: test your understanding
Try it yourself: Consider this case: Review one identity switch after two cars cross, even if every car was detected. Explain how you would check the result. In your answer, address this limitation: A good detection score does not guarantee stable tracks.
Check the reasoning
Tracking metrics include detection coverage and identity consistency, which measure different errors. A good detection score does not guarantee stable tracks.
What is the main idea behind evaluate tracks and identity switches?
Tracking metrics include detection coverage and identity consistency, which measure different errors.
How is it used in the example?
Review one identity switch after two cars cross, even if every car was detected.
What limitation should be checked?
A good detection score does not guarantee stable tracks.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 44 transcript313 words · Narration ready for the upcoming videoOpen transcript
Our parking tracker finds nearly every car box, but a red car changes identity twice while crossing the lot. A detection score might look excellent and still miss this problem.
Compare each predicted track with annotated identities over time. Count missed detections, false tracks, and identity switches, and inspect when they happen. An identity switch can matter greatly if the application counts individual visits or follows a vehicle between bays.
Show one clean sequence and one difficult crossing. Mark the frame where an identity changes, then trace whether the cause was a missed detection, an ambiguous association, or a long occlusion. A single aggregate score should be paired with these concrete errors. Detecting every car in every frame still does not guarantee good tracking. When two cars cross, their labels can swap while the boxes remain accurate. Review the crossing as a sequence, not a collection of isolated screenshots. Note where the identity changed, whether it later recovered, and how that affects the question the garage cares about. For occupancy, a brief swap may matter less than for a long-term vehicle count, so evaluate the error against the intended use. Now check the idea on a concrete example: Review one identity switch after two cars cross, even if every car was detected. A possible trap is worth naming: A good detection score does not guarantee stable tracks. This supports a careful conclusion: Tracking metrics include detection coverage and identity consistency, which measure different errors. When the red car changes labels, ask whether the garage's count changes too. The same switch may be serious for trip histories and harmless for a single occupancy snapshot. Report the identity error and the operational consequence separately.
Decide which errors affect the parking decision and set review rules accordingly. Next, we'll leave video tracking and represent a still image so we can search a collection for similar scenes.