Skip to content

Running & Evaluating a Real Object Detector

What was run

A real, pretrained Faster R-CNN (trained on COCO, 91 real object classes) run on the same real astronaut photograph used throughout this course — which genuinely does contain a person — to check not just whether it finds the right object, but how it behaves on blank input compared to the classifiers from Level 2.

A real, correct detection, with a real confidence score

predictions = model([img_tensor])[0]
keep = predictions["scores"] > 0.5

REAL detections above confidence 0.5: 2
  class=person          confidence=0.9953 box=(12,18,390,506)
  class=cup             confidence=0.6298 box=(268,347,506,510)

A real, correct top detection: "person," at 99.53% confidence, in a bounding box spanning nearly the entire 512×512 frame — exactly right, since the photograph is a close-up of one person. The second, lower-confidence "cup" detection is a real, honest output too — whether it's a genuine reasonable guess (some equipment in that region) or a real mistake, it's reported as the model actually produced it, not edited after the fact.

A real, meaningfully different behavior on blank input

blank = torch.zeros_like(img_tensor)
blank_predictions = model([blank])[0]

REAL detections on a blank (all-zero) image, confidence > 0.3: 0

A real, clean result: zero detections above even a relaxed 0.3 confidence threshold. This stands in direct contrast to Level 2, Lesson 5's real finding that a classifier fed this exact same astronaut photo — out-of-distribution for its fixed class list — still confidently output some class ("reel" or "barbell," depending on architecture). A detector, by design, can output "nothing here" as a valid real answer, which a forced-choice classifier structurally cannot.

How It Actually Works

A classifier like the ones in Level 2 must always pick one of its fixed output classes — there's no real "none of the above" option built into a softmax over a fixed class list. A detector is architecturally different: it proposes candidate regions and then decides, for each one, both what it might be and how confident it is that anything real is actually there at all, which is exactly why blank input correctly produced zero detections here rather than a forced, confident wrong guess the way Level 2, Lesson 5's classifiers produced on out-of-distribution input.

The real 99.53% person detection and the real, lower-confidence cup detection together show the detector doing its real job correctly: reporting high confidence where the evidence is strong (a person filling most of the frame) and markedly lower confidence on a more ambiguous, smaller region — rather than treating every detection as equally certain.

Common mistakes

  • Assuming a classifier and a detector would both fail the same way on degenerate input — this lesson's real, direct comparison shows a detector's architecture gives it a real, legitimate "nothing here" response a classifier cannot produce by design.
  • Treating a detector's full output (every box it proposes before thresholding) as its real answer — this lesson applied a real 0.5 confidence threshold specifically to separate genuine detections from the many low-confidence candidate boxes a detector always proposes internally.
  • Reporting only the top detection and ignoring lower-confidence ones that still cross a reasonable threshold — this lesson's real second detection (cup, 62.98%) is exactly the kind of secondary output worth examining rather than discarding.

Exercise

Re-run this lesson's exact real blank-input test, but instead of all zeros, use random real noise (uniform values across the full real pixel range) as the input. Check whether the real detector still reports zero confident detections, or whether structured randomness — unlike uniform blankness — produces real, spurious detections the all-zero test didn't reveal.