Lecture slides (a tiny correction on the non-max suppression slide made 2021.03.18.)
Weekly exercises (including a solution for the initial task)
If you are not able to access /projects/in5400/in5400_2020_lab_detection_data/, you can use this temporary link to a zipped copy of the data.
Hints for solving the additional questions:
- The immediate solution -- migh be computationally costly, though -- would be to apply your single-instance detector separately on a window covering (and some) each of the proposed bounding boxes in turn.
- Likely you would encounter that your detector would, for multiple of these proposal, end up targeting the same object. That is, you would have multiple detections of the same bird. To clean this up, so that you would ideally end up with a single detection, a single bounding box, for each bird, you could run the non-maximum suppression algorithm. [I am skipping explaining the algorithm here; see slides.]
- [I am not re-explaining the illustration here; use this more as a practice explaining this central concept.]
- The largest output boxes benefit from more "semantically rich" features. The FPN tries to propagate some of this strength back into the earlier layers used for detecting smaller objects.
Reading material