F-beta score for Detection API

This repo is a drop-in replacement for cocoapi to calculate F-beta scores for object detection tasks.

Installation

cd PythonAPI && make

OR

pip3 install 'git+https://github.com/yhsmiley/fdet-api.git#subdirectory=PythonAPI'

Usage

Initialize the api:

from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
cocoGt = COCO(true_path)  # initialize COCO ground truth api
cocoDt = cocoGt.loadRes(pred_path)  # initialize COCO prediction api
cocoEval = COCOeval(cocoGt, cocoDt, 'bbox')  # initialize COCO evaluation api
cocoEval.evaluate()

To get F-beta scores of detections:

cocoEval.accumulateFBeta()

Get best F-beta score for specific iou threshold and class (optional):

fscore, conf, precision, recall = cocoEval.getBestFBeta(beta=1, iouThr=0.5, classIdx=None, average='macro')

cocoEval.summarizeFBetaScores(average='macro')

cocoEval.printReport(beta=1, iouThr=0.5)

cocoEval.plotFBetaCurve(filename, betas=[1,2], iouThr=0.5, average='macro')

cocoEval.plotPRCurve(filename, average='macro')

Edge cases:

• no ground truths, and no detections for the class: TP=FP=FN=0 -> precision, recall, F1 = 1
• ground truth exists, but no detections for the class: TP=FP=0, FN>0 -> precision = 1

For original COCO-style evaluation:

cocoEval.accumulate()
cocoEval.summarize()
mapAll, map50 = cocoEval.stats[:2]

cocoEval.plotCocoPRCurve(filename, classIdx=None)

TODO

• Vectorize _getFBetaScore()
• Add option for micro/macro/weighted Fscores for _getFBetaScore()
• Combine similar functions
• Plot FBetaCurve per class