Recover from Occlusion and Missed Frames
A delivery van hides a parked car for two frames.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
A delivery van hides a parked car for two frames.
A practical question
A passing truck hides a small car for two frames. When the car reappears, should the system give it a new identity? A tracker can hold a predicted state briefly while detections are missing, but it must not pretend it observed the hidden car.
The useful idea
Set a time limit for retaining an unmatched track and define when a new detection may reconnect to it. Too short a limit fragments one trip; too long a limit may attach the old identity to a different car. Appearance and plausible movement can help, with uncertainty increasing during occlusion.
Work through the example
Review a clip where the truck crosses the scene. Mark observed boxes separately from predicted positions. When the car returns, inspect whether the same identity is justified by its path and appearance. A parked car disappears behind a passing truck and returns a moment later. If the system immediately deletes the old track, it may count a new car when the view clears. If it holds tracks forever, different cars can inherit the same identity. Use a short, documented tolerance for missing observations and check position, timing, and appearance when rejoining. Show the uncertain interval to a reviewer. An uninterrupted line on a display is not proof that every association was correct. Use this case to test your understanding: Flag uncertain identity until the hidden car becomes visible again. Here is the limitation to watch for: Keeping a track alive forever would create ghost cars. The answer follows from that evidence: A tracker can maintain a tentative state through short occlusions and reconnect later evidence. Replay the two missing frames separately from the later reappearance. The system should show that it predicted a location during occlusion rather than claiming it observed the car. Compare another car entering nearby before reconnecting the old identity.
Limits and next step
Keep the missed frames in evaluation. Next, we'll measure tracking quality across whole sequences, including identity changes that a frame-by-frame detector score misses.
Worked example
Flag uncertain identity until the hidden car becomes visible again.
Quick check: test your understanding
Try it yourself: Consider this case: Flag uncertain identity until the hidden car becomes visible again. Explain how you would check the result. In your answer, address this limitation: Keeping a track alive forever would create ghost cars.
Check the reasoning
A tracker can maintain a tentative state through short occlusions and reconnect later evidence. Keeping a track alive forever would create ghost cars.
What is the main idea behind recover from occlusion and missed frames?
A tracker can maintain a tentative state through short occlusions and reconnect later evidence.
How is it used in the example?
Flag uncertain identity until the hidden car becomes visible again.
What limitation should be checked?
Keeping a track alive forever would create ghost cars.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 43 transcript320 words · Narration ready for the upcoming videoOpen transcript
A passing truck hides a small car for two frames. When the car reappears, should the system give it a new identity? A tracker can hold a predicted state briefly while detections are missing, but it must not pretend it observed the hidden car.
Set a time limit for retaining an unmatched track and define when a new detection may reconnect to it. Too short a limit fragments one trip; too long a limit may attach the old identity to a different car. Appearance and plausible movement can help, with uncertainty increasing during occlusion.
Review a clip where the truck crosses the scene. Mark observed boxes separately from predicted positions. When the car returns, inspect whether the same identity is justified by its path and appearance. A parked car disappears behind a passing truck and returns a moment later. If the system immediately deletes the old track, it may count a new car when the view clears. If it holds tracks forever, different cars can inherit the same identity. Use a short, documented tolerance for missing observations and check position, timing, and appearance when rejoining. Show the uncertain interval to a reviewer. An uninterrupted line on a display is not proof that every association was correct. Use this case to test your understanding: Flag uncertain identity until the hidden car becomes visible again. Here is the limitation to watch for: Keeping a track alive forever would create ghost cars. The answer follows from that evidence: A tracker can maintain a tentative state through short occlusions and reconnect later evidence. Replay the two missing frames separately from the later reappearance. The system should show that it predicted a location during occlusion rather than claiming it observed the car. Compare another car entering nearby before reconnecting the old identity.
Keep the missed frames in evaluation. Next, we'll measure tracking quality across whole sequences, including identity changes that a frame-by-frame detector score misses.