Domino + OD — Scene Understanding (not just using largest bbox)

---

Created: 02 Feb 2023, 05:09 PM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a") Tags: knowledge,

---

Use Precision-Recall curve to get Average Precision, then mean across classes to get Mean Average Precision

• But run individually per image to get metric per image (using CocoDetection, run once per image)

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

https://cs230.stanford.edu/section/8/

https://towardsdatascience.com/what-is-average-precision-in-object-detection-localization-algorithms-and-how-to-calculate-it-3f330efe697b

But per image seems slow, 1.3s / image — ~1.5hrs for 5000 images

Can look at https://detectron2.readthedocs.io/en/latest/modules/evaluation.html?highlight=coco_evaluation#module-detectron2.evaluation instead to be faster

Results:

But when vary the type of AP used, e.g. using pred_AP_rangeIoU_largeArea

• Can have problems like the following:
• So should stick to pred_AP_rangeIoU_allArea