Represent Boxes and Confidence
A street photograph contains two parking signs that need separate boxes.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
A street photograph contains two parking signs that need separate boxes.
A practical question
A street image contains two signs. A classifier gives one label for the whole frame, but a detector can return a box for each sign with a class and a confidence score.
The useful idea
Boxes can be represented by corners or by a center, width, and height. Whatever convention we choose, document pixel scale and whether coordinates refer to the original or resized frame. A high confidence value ranks a detection under the model; it doesn't establish that the box is tightly aligned or that the sign is actually safe.
Work through the example
Overlay detections on the original street image. Check small signs, signs touching the border, and overlapping objects. Keep the annotation convention identical when comparing predictions with ground truth. A street image has two parking signs that require separate boxes. Draw their corner coordinates on the original and resized versions and verify the conversion by eye. A center-and-size representation can describe the same regions, but only if the receiver knows which convention was used. A high score ranks detections; it does not directly tell how many parking spaces are occupied. Test small signs and signs cut by the image edge before using the boxes in a count. Use this case to test your understanding: Draw each sign box in the original and resized image using one coordinate convention. Here is the limitation to watch for: A sign detection score is not a count of occupied parking spaces or a calibrated probability. The answer follows from that evidence: A box uses a defined coordinate convention; its score reflects model evidence for a class or object. The answer follows from that evidence: A detection score is not the same as per-space occupancy or a calibrated probability. Draw boxes for both signs before converting coordinates. If the resized-frame box is pasted onto the original image without undoing the scale, it lands in the wrong place. A visual overlay catches this format error faster than a plausible numeric score.
Limits and next step
One sign may receive several overlapping boxes. Next, we'll decide which duplicates to keep without accidentally removing two real nearby signs.
Worked example
Draw each sign box in the original and resized image using one coordinate convention.
Quick check: test your understanding
Try it yourself: Consider this case: Draw each sign box in the original and resized image using one coordinate convention. Explain how you would check the result. In your answer, address this limitation: A sign detection score is not a count of occupied parking spaces or a calibrated probability.
Check the reasoning
A box uses a defined coordinate convention; its score reflects model evidence for a class or object. A detection score is not the same as per-space occupancy or a calibrated probability.
What is the main idea behind represent boxes and confidence?
A box uses a defined coordinate convention; its score reflects model evidence for a class or object.
How is it used in the example?
Draw each sign box in the original and resized image using one coordinate convention.
What limitation should be checked?
A sign detection score is not a count of occupied parking spaces or a calibrated probability.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 53 transcript343 words · Narration ready for the upcoming videoOpen transcript
A street image contains two signs. A classifier gives one label for the whole frame, but a detector can return a box for each sign with a class and a confidence score.
Boxes can be represented by corners or by a center, width, and height. Whatever convention we choose, document pixel scale and whether coordinates refer to the original or resized frame. A high confidence value ranks a detection under the model; it doesn't establish that the box is tightly aligned or that the sign is actually safe.
Overlay detections on the original street image. Check small signs, signs touching the border, and overlapping objects. Keep the annotation convention identical when comparing predictions with ground truth. A street image has two parking signs that require separate boxes. Draw their corner coordinates on the original and resized versions and verify the conversion by eye. A center-and-size representation can describe the same regions, but only if the receiver knows which convention was used. A high score ranks detections; it does not directly tell how many parking spaces are occupied. Test small signs and signs cut by the image edge before using the boxes in a count. Use this case to test your understanding: Draw each sign box in the original and resized image using one coordinate convention. Here is the limitation to watch for: A sign detection score is not a count of occupied parking spaces or a calibrated probability. The answer follows from that evidence: A box uses a defined coordinate convention; its score reflects model evidence for a class or object. The answer follows from that evidence: A detection score is not the same as per-space occupancy or a calibrated probability. Draw boxes for both signs before converting coordinates. If the resized-frame box is pasted onto the original image without undoing the scale, it lands in the wrong place. A visual overlay catches this format error faster than a plausible numeric score.
One sign may receive several overlapping boxes. Next, we'll decide which duplicates to keep without accidentally removing two real nearby signs.