OBJECT DETECTION / OFFLINE TEACHING LAB
Which boxes should we return?
The same five parcel candidates as the lecture. Change one threshold at a time, then follow each NMS decision.
GREEDY SELECTION
| Candidate | Class | Score | IoU with A | State |
|---|
Evaluate the returned set
Dashed green is the single parcel annotation. Match by class and IoU, in score order, using each target at most once. This small example omits benchmark crowd and ignore rules.
About the data and thresholds
These are constructed teaching predictions, not output from a running model. The generated photograph and saved candidates are bundled in this file; there is no model download or network dependency.
Score filtering retains scores at least the cutoff. NMS suppresses a competing candidate only when its IoU is strictly greater than the NMS threshold. Class-aware NMS compares matching class labels only. The ambiguity example changes B's label without changing its score, box or colour.
Slides and lab are generated from results/detector_teaching_example.json using the same calculation functions. Data fingerprint: 2c1442e6ae9b47c4e07352a6360ddd4cd72cb990eb6317c6eee57533616b39a4