Diagnose Small and Occluded Objects
Distant and partly hidden signs are both missed for different reasons.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
Distant and partly hidden signs are both missed for different reasons.
A practical question
A sign far down the road occupies only a few pixels. Another sign sits behind a tree branch. The detector misses both, but for different reasons: the first has little recorded detail; the second has missing visible evidence.
The useful idea
Count errors separately by object size and degree of occlusion. Check whether resizing the input further discards small details, and whether the annotations include partly hidden signs consistently. Improving average recall on large clear signs may do nothing for these cases.
Work through the example
Compare original crops at their native resolution before enlarging them for display. A smoother enlargement cannot recreate text or boundaries the camera never sampled. For hidden signs, decide when an uncertain detection should go to a person instead of forcing a confident answer. A distant sign is small because the sensor recorded few pixels; enlarging its crop cannot restore missing detail. A nearby sign behind a branch is difficult for a different reason: part of it is not visible. Count the failures separately and inspect native-resolution crops. Check whether the annotation policy covers partly hidden signs consistently. An overall score that averages easy close signs with these cases can conceal the service's most consequential weakness. Now check the idea on a concrete example: Inspect native-resolution crops of small and occluded signs separately. A possible trap is worth naming: An overall average may conceal a failure on the cases the service needs. This supports a careful conclusion: Detection errors often depend on object size, occlusion, contrast, and annotation quality. Check the far sign at native resolution before blaming the detector. If the camera captured too little detail, an enlarged display will not restore it. For the sign behind branches, define when a human should review rather than forcing a confident box.
Limits and next step
Boxes tell us approximate locations. Next, we'll assign labels to individual pixels when the precise occupied area matters.
Worked example
Inspect native-resolution crops of small and occluded signs separately.
Quick check: test your understanding
Try it yourself: Consider this case: Inspect native-resolution crops of small and occluded signs separately. Explain how you would check the result. In your answer, address this limitation: An overall average may conceal a failure on the cases the service needs.
Check the reasoning
Detection errors often depend on object size, occlusion, contrast, and annotation quality. An overall average may conceal a failure on the cases the service needs.
What is the main idea behind diagnose small and occluded objects?
Detection errors often depend on object size, occlusion, contrast, and annotation quality.
How is it used in the example?
Inspect native-resolution crops of small and occluded signs separately.
What limitation should be checked?
An overall average may conceal a failure on the cases the service needs.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 56 transcript307 words · Narration ready for the upcoming videoOpen transcript
A sign far down the road occupies only a few pixels. Another sign sits behind a tree branch. The detector misses both, but for different reasons: the first has little recorded detail; the second has missing visible evidence.
Count errors separately by object size and degree of occlusion. Check whether resizing the input further discards small details, and whether the annotations include partly hidden signs consistently. Improving average recall on large clear signs may do nothing for these cases.
Compare original crops at their native resolution before enlarging them for display. A smoother enlargement cannot recreate text or boundaries the camera never sampled. For hidden signs, decide when an uncertain detection should go to a person instead of forcing a confident answer. A distant sign is small because the sensor recorded few pixels; enlarging its crop cannot restore missing detail. A nearby sign behind a branch is difficult for a different reason: part of it is not visible. Count the failures separately and inspect native-resolution crops. Check whether the annotation policy covers partly hidden signs consistently. An overall score that averages easy close signs with these cases can conceal the service's most consequential weakness. Now check the idea on a concrete example: Inspect native-resolution crops of small and occluded signs separately. A possible trap is worth naming: An overall average may conceal a failure on the cases the service needs. This supports a careful conclusion: Detection errors often depend on object size, occlusion, contrast, and annotation quality. Check the far sign at native resolution before blaming the detector. If the camera captured too little detail, an enlarged display will not restore it. For the sign behind branches, define when a human should review rather than forcing a confident box.
Boxes tell us approximate locations. Next, we'll assign labels to individual pixels when the precise occupied area matters.