mAP calculation (VOC 2012)

Refer to mAP implemention Cartucho/mAP, and support rboxes mAP eval here.

---

Quick start next.

folder

Creat directory as follow:

• detection-results: detections in standard format.
• ground-truth: gt in standard format.
• dets(op): your output, not recommended, just generate results in proper format in detection-results folder.
• gts: raw ground truth files.
• images-optional(op): test imgs, not recommended, cause it's really time-consuming.

format

• converter for each gt:
  <class_name> <x1> <y1> <x2> <y2> <x3> <y3> <x4> <y4> [--<difficult>]
  for each res(op):
  <class_name> <conf> <x1> <y1> <x2> <y2> <x3> <y3> <x4> <y4>
  and I'll release some common convert code here.

usage

the directory created above is root_dir, then you can start evaluation as follow"

from map import eval_mAP
eval_mAP(root_dir)

options

• difficult string attached to gt will escape certain object from being included.(such as HRSC, ICDAR).
• ignore variabel make it possible to ignore certain class from eval.
• Horizontal bbox eval can be conducted via coordinate trans, or just use this repo Cartucho/mAP.

map_func

make convenient for calling.