Timeline for Average Precision (AP) for object detection, huge confusion
Current License: CC BY-SA 4.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 19, 2023 at 14:42 | history | bounty ended | Tomé Silva | ||
| Jan 19, 2023 at 14:41 | vote | accept | Tomé Silva | ||
| Jan 18, 2023 at 19:29 | history | edited | Ciodar | CC BY-SA 4.0 |
clarified the preferred way to go
|
| Jan 18, 2023 at 18:55 | history | edited | Ciodar | CC BY-SA 4.0 |
added 1392 characters in body
|
| Jan 18, 2023 at 18:51 | comment | added | Ciodar | I've edited the answer. I think that the two approaches describe two different curves. The first one is sometimes called IoU- AP curve, while the second is the more common PR curve, whose integral yields the AP, and the mean of the per-class AP gives you the mAP. | |
| Jan 18, 2023 at 18:50 | history | edited | Ciodar | CC BY-SA 4.0 |
added 1392 characters in body
|
| Jan 18, 2023 at 15:22 | comment | added | Tomé Silva | If in fact they do the last thing I mentioned why not just compute a pair of precision-recall points per threshold, and integrate for AP? Instead of computing lots of precision recall points per threshold, doing AP of that and taking an average (of the AP's for every threshold) at the end ? | |
| Jan 18, 2023 at 15:16 | comment | added | Tomé Silva | Yes, you would keep IOU threshold fixed and compute the precision and recall for the entire evaluation set. You would have as many pairs (precision, recall) as thresholds. That would give you a graph at which you can integrate to get average precision. (cocodataset.org/#detection-eval). At least is what I understand from their description. But it could be that they do the procedure you just described for every threshold and then just average everything at the end. | |
| Jan 18, 2023 at 14:02 | comment | added | Ciodar | Could you provide the source where you read about these two approaches? It is not clear in the first approach if you intend to keep the IOU threshold fixed for each computation of precision and recall of the detections (which seems logic to me) or not. | |
| Jan 18, 2023 at 13:50 | comment | added | Tomé Silva | Thank you for your answer. So the relevance of this metric is that it shows precision across the confidence level of the various detections. I think this is what I was missing. The first approach, where you change the threshold won't reflect the effects of predictions confidence. My question would be then why use one or the other? | |
| Jan 17, 2023 at 22:35 | history | answered | Ciodar | CC BY-SA 4.0 |