OD

---

Created: 27 Jan 2023, 03:00 PM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a") Tags: knowledge,

---

https://jonathan-hui.medium.com/map-mean-average-precision-for-object-detection-45c121a31173

Confidence score threshold will affect whether the model even identifies the box as a class before this process.

Actual confidence score is not impt in identifying TP/FP/FN?

For each class in classes:

Get gt box indexes in that class (and the len)

Get dt box indexes in that class (and the len)

⇒ Get precision/recall for each class for 1 image

• If photo has 1 gt, 1 dt and is a TP, precision = 1, recall = 1

⇒ average across all classes (including those not present (dt/gt) in this image?) for mAP?

• If do this ⇒ if the photo itself only has 1 gt, 1 dt and is a TP, how to ensure mAP is high?

For 100 detection boxes (15 unique labels) and 20 ground truth boxes (10 unique labels) ⇒ 17 unique labels

For each 17 unique labels:

Problem — average precision is