Computer Vision with Neural Networks
Keep distinct seed detections
A removed middle box overlaps a later box
Botie has overlapping seed detections with different scores. Greedy suppression must keep distinct candidates without allowing an already removed box to remove another one.
BOTIE
The inputs we are comparing
Candidate seed rectangles
One duplicate and one separate seed
| 0 | 0 | 4 | 4 |
| 0.5 | 0 | 4.5 | 4 |
| 7 | 0 | 9 | 2 |
Removed boxes cannot suppress
| 0 | 0 | 3 | 2 |
| 1 | 0 | 4 | 2 |
| 2 | 0 | 5 | 2 |
Candidate scores
Same in both cases
- 00.9
- 10.8
- 20.7
IoU suppression threshold
One duplicate and one separate seed
0.5
Removed boxes cannot suppress
0.4
Result for one duplicate and one separate seed
| 0 | 0 | 4 | 4 |
| 7 | 0 | 9 | 2 |
The strongest seed candidate arrives last in the list
This check uses the pieces from “Rank seed candidates before suppressing overlaps”.
BOTIE
- Candidate seed rectangles
0 0 4 4 6 0 8 2 0.2 0.2 4.2 4.2 - Candidate scores
- 00.6
- 10.7
- 20.95
- IoU suppression threshold
0.5
Use a number, text, or an array matching the requested result. For a matrix, put each row in brackets. Saving your prediction runs your current construction on these inputs and checks the earlier examples too.
Build a construction in the workspace above first. This case will test those pieces.
What should happen when IoU equals the threshold exactly, and why might suppression still remove a real nearby object?
Greedy suppression is a heuristic with an explicit equality rule and a tradeoff between duplicates and nearby instances.
All candidates here belong to one class. Ties use the original index; real library tie behaviour can differ. Suppression can mistakenly remove nearby real objects.
Work through the Python challenges →