Why do you use topK instead of a fixed threshold for output? Wouldn't it make a big error if there were only a few or a lot of targets in the image?
bruceyang2012
All 2 comments
I use top 100 detections for COCO dataset. The COCO evaluation metric does not give any penalty for low confident false prediction. In a real application, you can set a hard threshold, e.g., 0.3. It will kill most false positives. I haven't tried set hard thresholds for COCO evaluation, it is for sure lower AP than top 100. It is easy to try it yourself.
xingyizhou
on 24 Apr 2019
👍4
@xingyizhou I see, thank you.
bruceyang2012
on 24 Apr 2019
Was this page helpful?
0 / 5 - 0 ratings
Related issues
lih627
路
7Comments
VRCMF
路
6Comments
HamsterHuey
路
4Comments
pipi2020
路
7Comments
bhack
路
6Comments
Most helpful comment
I use top 100 detections for COCO dataset. The COCO evaluation metric does not give any penalty for low confident false prediction. In a real application, you can set a hard threshold, e.g., 0.3. It will kill most false positives. I haven't tried set hard thresholds for COCO evaluation, it is for sure lower AP than top 100. It is easy to try it yourself.