Skip to content
Free public beta until December 4, 2026: every course is open, no card needed. After that, the first chapter of each course stays free and full access is a one-time $39 ($19 for up to 100 founding members chosen for helpful feedback). Your progress is kept. How the beta works

Recipe

Run the object detection pipeline end to end

Use when: you have a trained detector and a new image, and need a final, clean list of detected objects

Goal: a set of final detections, each with a class, a box, and a confidence score, cleaned of duplicates and thresholded to a usable confidence level

Taught in Precision, recall, and mAP

Quick recipe

  1. Preprocess the image
  2. Run the detector's backbone and head
  3. Collect the raw boxes and scores
  4. Filter by confidence threshold
  5. Apply non-max suppression
  6. Report the final detections
  1. 1

    Resize, normalize, and format the image the way the detector's own input pipeline expects (Chapter 12's own preprocessing conventions).

    Why: a detector's backbone was trained on inputs in one specific format. Feeding it anything else produces meaningless output, not an error.

  2. 2

    Pass the preprocessed image through the network: a backbone (Chapter 12) extracts features, and a detection head turns those features into raw candidate boxes with class scores.

    Why: this is the network doing the actual "where might objects be" work. Everything after this step is bookkeeping on its output, not more learning.

  3. 3

    Read off every candidate box the head produced, each with a class label and a confidence score (Section 1). For a one-stage detector, this is usually hundreds of boxes, most of them near-duplicates or noise.

    Example

    A detector reports 340 raw boxes for one image. Most are low-confidence noise or duplicates of the 5-6 real objects actually present.

    Why: nothing has been filtered yet. This is the pipeline's rawest, most honest checkpoint, useful for debugging (Lab B).

  4. 4

    Discard every box whose score falls below a chosen threshold (Section 4's own worked example: 0.5, 0.6, 0.9 all give different, valid results).

    Why: a low-confidence box is the detector's own way of admitting it isn't sure. Keeping every one of them would flood the final result with noise.

  5. 5

    Run NMS (Section 2) on the boxes that survived filtering, separately per class, to collapse duplicate detections of the same object down to one.

    Why: without this step, one real object gets listed many times. That wrecks precision, even though every individual box might be "correct."

  6. 6

    The boxes left after both filters are the pipeline's actual output. Report each one's class, box coordinates, and confidence score.

    Why: this is the output everything upstream exists to produce: a clean list of detections, with duplicates removed and confidence high enough to trust.

Common mistakes

  • Running NMS before the confidence filter, instead of after: this wastes the suppression step comparing boxes that a simple threshold would have removed for free.
  • Running NMS across all classes at once instead of per class: this wrongly removes one of two genuinely different objects that happen to overlap (Section 2).
  • Treating raw, not-yet-filtered detector output as "the result": this is Step 3's checkpoint, not the pipeline's actual output.
  • Comparing two detectors' mAP scores from different datasets, or different IoU conventions, as if they were the same measurement.